A Comparative Ablation Study of CNN-LSTM-GRU and MAformer Architectures for Operational Multi-regime Salinity Forecasting in the Outer Shannon Estuary

A Comparative Ablation Study of CNN-LSTM-GRU and MAformer Architectures for Operational Multi-regime Salinity Forecasting in the Outer Shannon Estuary | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Comparative Ablation Study of CNN-LSTM-GRU and MAformer Architectures for Operational Multi-regime Salinity Forecasting in the Outer Shannon Estuary Opeyemi Ajibola-James This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9178680/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Accurate Sea Surface Salinity (SSS) forecasting is critical for operational management in macrotidal environments, yet the transition between local tidal forcing and long-term climatological drivers remains poorly understood in deep learning. This study presents a multi-regime evaluation and comparative ablation analysis of a CNN-LSTM-GRU (Hybrid) model versus a Multi-Head Attention Transformer (MAformer) in the Outer Shannon Estuary. Leveraging nine robust predictors categorized into Hydrodynamic, Atmospheric, and Steric/Thermal groups, an "endurance test" across horizons from 24 to 240 hours was conducted. To ensure physical consistency, a scaled lookback strategy (L = h + 1) was implemented, providing models with up to 19.4 M2 tidal cycles of historical context. Results demonstrate a distinct architectural crossover. The Hybrid model provides superior short- and medium-range stability (24h R 2 = 0.892 and 120h R 2 = 0.518), yet reaches architectural de-coherence at 240 hours, characterized by a performance collapse (R 2 = 0.206). Conversely, the MAformer exhibited superior long-term resilience, achieving lower error magnitudes (RMSE: 0.405 vs. 0.438) at the 10-day horizon. The thematic ablation reveals a scale-dependent regime shift: short-term forecasts are dominated by local velocity signals, whereas 240h stability is entirely dependent on global Steric/Mass and Thermal "Climatological Anchors". Without these anchors at 240h, both architectures experience total predictive failure; notably, the MAformer’s R 2 dropped from ~ 0.216 to 0.003. Findings suggest that operational estuarine systems should adopt a hierarchical modelling approach: deploying Hybrid units for daily navigational safety and Attention-based architectures for long-term strategic planning to maintain physical-mathematical alignment across expanding temporal scales. Artificial Intelligence and Machine Learning Physical Geography Oceanography Hydrology Outer Shannon Estuary SSS Forecasting MAformer vs CNN-LSTM-GRU Ablation Study Architectural De-coherence Multi-Regime Modelling Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Estuaries represent some of the most complex and dynamic aquatic environments on Earth. As transitional zones between fluvial freshwater systems and the saline open ocean, they are characterized by intense gradients in salinity, temperature, and nutrient concentration (Pritchard, 1967). SSS is perhaps the most critical parameter in these systems, serving as a primary indicator of water density, circulation patterns, and habitat suitability for diverse marine species. Accurate SSS forecasting is essential for operational decisions in desalination, maritime navigation, and the management of sensitive aquaculture sites. Historically, estuarine modelling has relied on numerical hydrodynamic models, such as the Finite-Volume Community Ocean Model (FVCOM) or the Regional Ocean Modelling System (ROMS). While these models are physically consistent, they are computationally expensive and often struggle with the "curse of dimensionality" when incorporating high-frequency local sensor data alongside global satellite products. Recent advancements in computational hydrology have shifted toward the application of deep learning (DL) architectures to capture the non-linear and multi-scale temporal dependencies inherent in estuarine salinity dynamics. Models like Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRU) have demonstrated an exceptional ability to capture the temporal dependencies of tidal cycles (Hochreiter & Schmidhuber, 1997; Cho et al., 2014). The LSTM networks have emerged as a foundational tool due to their ability to mitigate the vanishing gradient problem in long-term environmental time series. Gorski et al. (2024) demonstrated that LSTM-based models could achieve the physical accuracy of traditional 3D hydrodynamic models in the Chesapeake and Delaware Bays while operating at a fraction of the computational cost. Similarly, Saccotelli et al. (2024) validated the efficacy of LSTMs in predicting salt wedge intrusions in the Po River Delta, highlighting their superior performance in micro-tidal environments compared to traditional machine learning algorithms. To further enhance predictive stability, researchers have begun integrating hybrid DL structures and meta-heuristic optimization algorithms. Qi et al. (2023) expanded this by introducing residual connections (Res-LSTM and Res-GRU) combined with transfer learning to address data scarcity and improve generalizability across different estuarine regimes. Nguyen et al. (2025) proposed a hybrid LSTM-GRU model optimized with the Sailfish Optimizer (SFO) for the Hau Estuary, reporting significant improvements in 7-day forecast horizons. The integration of GRU is often preferred for its reduced parameter complexity without sacrificing the performance of gated mechanisms. The trend toward "Physics-Informed" or "Hybrid-Mechanism" modelling is also gaining traction. Optimization of hyperparameters remains a critical focal point; Zheng et al. (2024) utilized an Improved Sparrow Search Algorithm (ISSA) to tune LSTM parameters in the Qiantang Estuary, yielding about 10% reduction in RMSE. Complementary to these temporal models, Zhu et al. (2024) emphasized the role of data preprocessing, showing that Discrete Wavelet Transform (DWT) decomposition within an Artificial Neural Network (ANN) framework significantly sharpens the model’s sensitivity to extreme high-salinity events triggered by sea-level rise. Zhang et al. (2025) coupled DL models with the FVCOM in the Changjiang Estuary, finding that the DL component reduced numerical simulation errors by over 50% by learning the residual biases of the physical model. Despite these advances, some significant research gaps remain. First, the European coastline and the allied outer estuaries that are characterized by complex tidal regimes and diverse freshwater discharge patterns (particularly including the outer parts of Shannon, and Tagus estuaries) are still understudied with tripartite variants of such hybrid data-driven models, particularly CNN-LSTM-GRU; and evolving generalist models for time series forecasting, particularly MAformer. Second, most DL salinity models still focus on short-term horizons (e.g., 24-48 hours), and a few focus on medium-term horizons (e.g., 72–120 hours). There is a dearth of research into how the predictive power of DL architectures decays over "Endurance Tests" extending to 10 days (240 hours), a timeframe critical for long-term water resource planning. Third, the relative importance of environmental predictors is often assumed to be static. However, estuarine physics suggests that the drivers of a 24-hour forecast (dominated by M2 tides) differ fundamentally from the drivers of a 240-hour forecast (influenced by Spring-Neap transitions and global steric sea-level changes). This suggests lack of scale-dependent ablation. There is no rigorous evidence in current literature that demonstrates how the "information value" of a specific predictor (e.g., Steric vs. Thermal) shifts as the forecast moves from a tidal regime (24h) to a climatological regime (240h). The gaps present apparent need for a rigorous comparative study of a Hybrid CNN-LSTM-GRU architecture and a MAformer for an extended multi-horizon (24-240 hours) SSS forecasting in a relatively complex outer part of European estuary. This should involve a systematic ablation study across five horizons that could decode the "information value" of thermal, steric, surface, and velocity predictors. Additionally, such study should address the gap of "Information Starvation" by scaling the lookback window to capture up to 19.4 tidal cycles for long-term horizons, a methodology not yet standardized in European estuarine DL research. In this regard, the objectives of this study are to: (i) evaluate the performance decay of Hybrid and MAformer architectures across 1-day (24-h) to 10-day (240-h) forecast horizons; (ii) identify the scale-dependent importance of environmental predictors through thematic ablation; and (iii) establish an operational hierarchy for sensor-efficient estuarine monitoring based on forecast lead times. 2. Study Area The Shannon Estuary, the largest estuarine system in Ireland, extends approximately 100 km from Limerick City westward to the Atlantic Ocean between Loop Head (County Clare) and Kerry Head (County Kerry), encompassing a navigable surface area of roughly 500 km². It forms the tidal outlet of the River Shannon—Ireland’s longest river—and represents a classic macrotidal environment characterized by strong tidal forcing, complex bathymetry, and pronounced salinity gradients. The estuary exhibits highly dynamic hydrographic conditions, where rapid fluctuations in sea surface salinity (SSS) are a defining physical feature and an important ecological control (Costello, 2025). The monitoring (observations) point used in this study is situated in the outer estuary (Figure 1), where marine–fluvial interactions are particularly intense. The outer part of the Shannon Estuary is characterised by significant spatial and temporal dynamics, which underscore a complex hydrographic regime characterised by distinct geographic gradients and pronounced seasonal transitions (Ajibola-James, 2026). Points A, B, C, and D have annual mean SSS (AMSSS) of 33.985, 33.881, 34.125, and 34.343; and annual mean CV (AMCV) of 0.086, 0.073, 0.094, and 0.106 % respectively (Figure 1). The lowest values exhibited by the point B (52.500000 o N, -9.74999 o W), the data observations point imply the highest level of annual freshwater availability (AFWA); and the most stable SSS on annual time scale (Ajibola-James, 2026). Salinity variability within the Shannon Estuary plays a critical role in structuring biological processes and ecosystem functioning. In particular, marked SSS oscillations in the inner estuary have been shown to influence lipid and fatty acid composition in key estuarine biota, highlighting salinity as a primary environmental stressor (Costello, 2025). Beyond its ecological relevance, the estuary and its catchment have substantial socio-economic importance, especially for Ireland’s Mid-West Region. The River Shannon supports navigation, fisheries, industry, and potable water supply; however, this resource is vulnerable to upstream seawater intrusion (USI), which poses risks to freshwater abstraction and water security under changing hydrodynamic and climatic conditions. 3. Methodology 3.1 Datasets and the Observation Point The most appropriate data observation point in the outer Shannon Estuary for this study was statistically determined with the most relevant recent finding on the SSS variability by Ajibola-James ( 2026 ) that established the point B as the most appropriate for monitoring and predicting USI in the outer estuary. Additionally, the need for making this a foremost task in this SSS prediction study is in consonance with his opinion. The study utilised datasets collected from the point B, incorporating SSS and 11 potential predictors (01 Mar. 2024, 00:00 hr-30 Jun. 2025, 23:00 hr), Global mean sea-level variations (gmsv), Global mean mass volume variation (gmmvv) that is also called Global average sea level change due to water mass; Sea water potential temperature (swpt); Inverted barometer (ib) that depicts Change in sea surface height due to change in air pressure; the allied wind speed components, particularly Sea surface water stokes drift x velocity (vsdx) and Sea surface water stokes drift y velocity (vsdy); sea surface height above geoid (ssh); Northward sea water velocity (nswv); Eastward sea water velocity (eswv); Sea surface height above geoid due to ocean tide (otide); and Surface sea water y velocity due to tide (vtide). The datasets were retrieved from the online data repository of the European Union (EU) Copernicus Marine Service Information (CMEMS) using the Global Ocean Physics Analysis and Forecast (GOPAF) product, L4 (hourly). 3.2 Data Cleaning: Data Quality Control and Anomaly Remediation Preliminary inspection of the 16-month datasets at hourly time scale revealed that missing values (NAs) were minimal (approx. 3.06%); and 2 predictor variables, specifically the velocity components (vsdx and vsdy) exhibited 'hard zero' anomalies. These were defined as continuous sequences of 0.000 values exceeding two hours—a state physically inconsistent with the dynamic tidal regime of the Shannon Estuary. The order of the anomaly remediation was characterized by zero-to-NA followed by hybrid imputation. To ensure temporal continuity for the deep learning architectures, a hybrid imputation strategy was implemented using the imputeTS framework in R software. The pacman , and tidyverse libraries in R were also utilised during the data cleaning tasks. Isolated missing points and short-duration gaps (less than 10 hours) were remediated using linear interpolation. In contrast, larger gaps (above 10, reaching up to 30 hours in the 2 variables) were reconstructed using State-Space Kalman Smoothing (based on a structural time-series model). This approach preserves the underlying sinusoidal tidal harmonics, ensuring that the CNN-LSTM-GRU filters and MAformer attention heads receive a physically coherent signal rather than artificial step-functions or flat-line artifacts. This was achieved prior to the implementation of the three-stage variable selection pipeline with appropriate R libraries to ensure that the features selection was driven by physical dependencies rather than sensor artifacts. Technically, this ensured that the information-theoretic metrics (Mutual Information) and recursive elimination (RFE) operated on a continuous and physically representative state-space. 3.3 Multicollinearity Detection To ensure that this study meets the highest standards of statistical rigor and hydrological/oceanographic operations, a Variance Inflation Factor (VIF) analysis with appropriate R libraries ( pacman, car, tidyverse ) coupled with the author’s Hydrologic/Oceanographic domain intelligence were leveraged to detect the inherent multicollinearity in the 11 potential SSS predictors. VIF (Multivariate) measures how much the variance of a regression coefficient is inflated due to collinearity with all other predictors in the model. It detects "collective" overlap that a simple correlation matrix (e.g. Pairwise Correlation) might miss. The VIF decision rules usually involve removal of high multicollinearity variables with VIF > 10; and retention of the moderate multicollinearity variables (VIF > 5 but < 10); and relatively low multicollinearity variables (VIF < 5). The initial multicollinearity assessment via the VIF indicated that all the 11 potential predictors remained below the critical threshold of 10, with gmsv (7.8767) exhibiting the highest collinearity. Notably, astronomical tidal height (otide) and tidal velocity (vtide) displayed VIF values near 1.01, suggesting statistical independence (Fig. 2 ). In the context of the Shannon Estuary, variables like different velocity components or overlapping pressure indices often move so closely together that they can easily mask each other's true importance (Zhang et al., 2021), particular including in VIF analysis. Consequently, the 2 predictors with the relatively moderate VIF, gmsv (7.8767) and gmmvv (4.9714) that could have been erroneously deleted by the researcher to achieve the goal of retaining 9 robust predictors for the deep learning modelling were retained. The decision to retain them till the time of implementing the relevant 3-stage feature selection pipeline was based on the author’s Hydrologic/Oceanographic domain intelligence as subsequently detailed in 3.4. 3.4 Feature Selection (The 3-stage Pipeline) The utilised 3-stage feature selection pipeline was implemented with relevant R libraries ( pacman, praznik , caret , xgboost , tidyverse , scales , and randomForest ). The final predictor suite was determined by a consensus ranking of three distinct feature selection paradigms: (a) Mutual Information, representing non-linear statistical dependence; (b) RFE, representing optimal subset performance via 10-fold cross-validation; and (b) XGBoost Gain, representing decision-tree-based predictive contribution. These three different mathematical perspectives are usually depicted by a single defensible truth called Global Score (GS). High GS (e.g., > 0.8) indicates that the variable performed exceptionally well across all three tests. This is the universal importance index, which implies that the variable is a "Primary Driver" of salinity. Moderate GS (e.g., 0.4–0.7) indicates that the variable might be very strong in non-linear relationships (High MI) but slightly less efficient at tree-splitting (Lower XGB). This is the conditional importance index, which implies that the variable is a "Secondary Driver". Low GS (e.g., < 0.3) indicates that the variable provides some value but is likely on the verge of being "noise" rather than "signal". This is the marginal importance index, which implies that the variable is a “Minor Driver”. As shown in Table 1 , the 'Robust 9' predictors (green with Rank 1–9) demonstrated high convergence across all three metrics. Notably, the 2 variables with the relatively high but moderate multicollinearity, gmsv (VIF = 7.8767) and gmmvv (VIF = 4.9714) were objectively retained only because their relatively high GS (0.9696, and 0.6594 respectively) ranked within the top 9. Conversely, the otide (VIF = 1.0113) and vtide (VIF = 1.0082) that showed the least multicollinearity were automatically eliminated due to their relatively low GS rank of 0.0099 and 0.0037 (yellow with Rank 10–11) respectively. The use of the GS rank as the final assessment index was to ensure that the selected feature space prioritized predictive power while maintaining numerical stability for the MAformer and CNN-LSTM-GRU models. It should be underscored that otide and vtide are mathematical constituents (harmonics), not physical "drivers" like wind stress or potential temperature. The automatic exclusion of the 2 potential predictors is also in alignment with the "Thematic Ablation" Framework that is developed and utilized in this study, which grouped the SSS predictors into Steric/Mass, Thermal, Surface, and Local Velocity. Scientifically and realistically, to maintain the integrity of a Thematic Ablation Study, the predictors must represent distinct physical processes in the real-world. Astronomical tides are a constant background oscillation; the variability in SSS—which is the target variable—is driven by the 9 predictors that are objectively retained. The intricate tidal-fluvial interaction within the Shannon makes velocity components particularly prone to mutual dependency, requiring careful feature selection to reveal their distinct roles in salinity advection (O'Donncha et al., 2020 ). The exclusion of the 2 predictors focuses the architectures on the primary physical drivers of salinity transport—Steric, Thermal, Surface, and Velocity regimes—while preventing the Attention mechanism from over-fitting to theoretical harmonics. Thus, running such a VIF in accordance with appropriate domain knowledge in Hydrology/Physical Oceanography before the subsequent RFE and XGBoost stages helps to clean the field of redundant predictor variables. Table 1 Rank of the SSS potential predictors based on the Global Score. Rank Predictor Variable MI Score AvgImportance RFE Norm XGB Gain MI Norm XGB Norm Global Score 1 gmsv 0.624417 54.83446 0.908815 0.622876 1.000000 1.000000 0.969605 2 gmmvv 0.435738 58.57355 1.000000 0.179921 0.689391 0.288662 0.659351 3 swpt 0.444438 48.20372 0.747112 0.172262 0.703714 0.276363 0.57573 4 ib 0.086757 43.54378 0.633471 0.010145 0.114891 0.01602 0.254794 5 vsdx 0.086377 36.42265 0.459808 0.008185 0.114264 0.012873 0.195649 6 ssh 0.037979 37.95439 0.497163 0.002237 0.034591 0.003321 0.178358 7 vsdy 0.032835 34.64113 0.416363 0.001991 0.026122 0.002926 0.14847 8 nswv 0.019623 25.51207 0.193733 0.001169 0.004372 0.001607 0.066571 9 eswv 0.017934 25.13565 0.184554 0.000788 0.001593 0.000995 0.062381 10 otide 0.016967 18.78234 0.029616 0.000258 0.000000 0.000144 0.00992 11 vtide 0.023656 17.56791 0.000000 0.000169 0.011012 0.000000 0.003671 3.5 Conceptual Framework: Endurance Test Between CNN-LSTM-GRU and MAformer The first part of the modelling study was focused on endurance test between CNN-LSTM-GRU and MAformer, which leveraged the 9 robust predictors for multi-regime SSS forecasting (24, 72, 120, 168, and 240 hours ahead). The appropriate R libraries, pacman , tidyverse , keras3 , tensorflow , Metrics , ggplot2 , scales , and reshape2 were leveraged to achieve this part. The utilised conceptual framework serves as the "logical blueprint" of this part of the research by bridging the gap between environmental physics (the predictors), computational engineering (the architectures), and operational decision-making (the horizons). The framework was built on a "Multiscale Predictive Logic" that explains why the model's structure must change as the forecast horizon expands. It is structured into four integrated modules, (a) Environmental Forcing, (b) Temporal Scaling, (c) Architectural Processing, and (d) Operational Evaluation as subsequently detailed. This framework is designed to move beyond simple "black-box" forecasting by establishing a physically grounded relationship between input lookback and output lead time. (a) Environmental Forcing Module (The Inputs) The framework begins by categorizing the 9 robust predictors into three physical "influence layers" that drive salinity flux in the Outer Shannon Estuary: Layer 1 (The Local Pulse): High-frequency hydrodynamic signals consisting of local seawater velocity (nswv, eswv) and total sea surface height (ssh). Layer 2 (The Surface Forcing): Synoptic atmospheric variables including wind stress components (vsdx, vsdy) and inverted barometer pressure (ib). Layer 3 (The Climatological Anchor): Deep-memory variables including seawater potential temperature (swpt) and global steric/mass variations (gmsv, gmmvv). (b) Temporal Scaling Module (The Logic) To capture the cyclic nature of the estuary, the framework applied a physics-proportional windowing/scaling for the lookback window (L) relative to the forecast horizon (h) such that L = h + 1. For example, for the 24-hour study, L = 25 (capturing 2 tidal cycles), while for the 240-hour study, L = 241 (capturing ~ 19.4 tidal cycles). This logic is critical for the "Endurance Test," as it provides the architecture with approximately 19.4 M2 tidal cycles, allowing the model to "see" the transition between Spring and Neap tidal phases before making a relatively long-term SSS prediction (10 days in the future). (c) Architectural Processing Module (The Engines) The two distinct computational philosophies contrasted in the study are as follows: Hybrid (CNN-LSTM-GRU): It focuses on Spatio-Temporal Feature Extraction. The CNN layers filter local spatial gradients, while the LSTM and GRU units handle short-to-medium range temporal "memory". This architecture consists of a 1D-Convolutional layer (64 filters, kernel size 3) for local pattern recognition, followed by an LSTM layer (100 units) and a GRU layer (50 units) to capture hierarchical temporal features. A dropout rate of 0.2 was applied to mitigate overfitting. MAformer (Attention-based): Focuses on Long-Range Dependency Mapping. The Multi-Head Attention mechanism identifies non-linear correlations across the entire 241-hour sequence, bypassing the vanishing gradient limitations of traditional recurrent units. The MAformer employs a Multi-Head Attention mechanism with 4 heads and a key dimension of 25. This allows the model to simultaneously attend to different periodicities (e.g., the 12.42h tidal cycle and the 24h solar cycle). The architecture includes Layer Normalization and Feed-Forward Networks to stabilize the learning process. (d) Operational Evaluation Module (The Regimes) The final stage of the framework groups the five horizons in Table 2 into three operational regimes, allowing for a strategic interpretation of the results, short-term (24h) that focused on navigational safety and tidal advection; the medium-term (72h–120h) that focused on operational dredging and synoptic weather planning; and long-term (168h–240h) that focused on strategic habitat management and climatological drift. Table 2 Operational regimes and configuration for the multi-regime SSS forecasting. Operational Regime Forecast Horizon (h) Optimal Lookback (L) Tidal Cycles (approx. M2​) Physical Significance & Scaling Logic Short-term 24 Hours 25 Hours ~ 2.0 Cycles Resolves daily semidiurnal rhythms and immediate tidal advection. Medium-term 72 Hours 73 Hours ~ 5.9 Cycles Captures synoptic weather shifts and cumulative tidal phase drift. 120 Hours 121 Hours ~ 9.7 Cycles Balances high-frequency data with 5-day estuarine "memory." Long-term 168 Hours 169 Hours ~ 13.6 Cycles Captures nearly a full 14-day Spring-Neap tidal transition. 240 Hours 241 Hours ~ 19.4 Cycles Resolves climatological trends and full Spring-Neap cycles. 3.6 Standardized Training Protocol and Error Metrics To ensure a fair comparison across all horizons and ablation groups, all the models utilised in the first and second parts of the modelling study were trained with: Epochs: 20 Patience: 12 (Early Stopping) Optimizer: Adam Loss Function: Mean-Squared-Error (MSE) Similarly, the accuracy of all the models was assessed with the coefficient of determination (R 2 ); and the absolute error metrics, root-mean-squared-error (RMSE) and mean-squared-error (MAE). The higher the value of R 2 , the better a model’s performance in terms of variation explained; the lower the RMSE and MAE, the better a model’s forecasts accuracy. 3.7 Conceptual Framework: Ablation Test Between CNN-LSTM-GRU and MAformer The second part of the modelling study was focused on ablation test between CNN-LSTM-GRU and MAformer, which also leveraged the same 9 objectively-selected predictors for the baseline, multi-regime SSS forecasting (24, 72, 120, 168, and 240 hours ahead). The suitable R libraries, pacman , tidyverse , keras3 , tensorflow , Metrics , ggplot2 , scales , and reshape2 were leveraged to achieve this part. Figure 3 presents the thematic ablation workflow diagram. Generally, a credible ablation study helps to "unmask" the predictors importance that simple correlation matrices (e.g. Pairwise Correlation) might miss. The ablation study reveals how predictor importance shifts across the horizons. The 9 robust predictors were categorized into four thematic groups namely (a) Steric/Mass (gmsv, and gmmvv), (b) Thermal (swpt), (c) Surface (ib, vsdx, vsdy, and ssh), and (d) Local Velocity (nswv, and eswv) for the ablation study. In effect, the ablation group for the multi-regime SSS forecasting consists of Baseline (Full), No Steric/Mass, No Thermal, No Surface, and No Local Velocity. To make the results of the ablation test comparable for supporting relevant operational decisions, the same “Conceptual Framework” in 3.5; and the same “Standardized Training Protocol” in 3.6 were utilised. 4. Results and Discussion 4.1 Comparative Endurance Analysis (Hybrid Vs MAformer) Table 3 shows the performance of the CNN-LSTM-GRU (Hybrid) and MAformer architectures across 5 forecast horizons. Their performance in terms of Goodness of Fit and horizon decay is also presented in Appendices A and B. A distinct "performance crossover" was observed as the lead time extended from the inception of the long-term (168 h) to the extreme of the long-term (240 h) regime. In the short-term regime (24h), the Hybrid model demonstrated superior precision, achieving an R 2 of about 0.892 compared to the MAformer’s R 2 of about 0.751. This suggests that for immediate tidal advection, the CNN’s local feature extraction combined with the GRU’s recurrent memory is more efficient at mapping high-frequency fluctuations. At the 120-hour horizon, the Hybrid model demonstrates its maximum performance advantage, achieving an RMSE that is 22.3% lower than that of the MAformer (0.333 vs. 0.429 respectively). This significant 0.0955 error gap underscores the Hybrid's superior ability to maintain phase alignment during the critical synoptic transition phase (Appendix B)—a finding consistent with Schmidt et al. ( 2021 ), who noted that RNN-based hybrids excel when the local hydrodynamic "memory" is still physically relevant. However, as the horizon reached the extreme of the long-term regime (240h), the Hybrid model’s R 2 decayed significantly to about 0.205 while the RMSE show unprecedented increase to about 0.438. In contrast, the MAformer exhibited higher resilience, outperforming the Hybrid with an R 2 of about 0.241 and a lower RMSE of about 0.405. The MAformer’s advantage at 10 days is attributed to its self-attention mechanism, which effectively processed the 241-hour lookback window. While the recurrent units in the Hybrid model likely suffered from information dilution over ~ 19.4 M2 tidal cycles, the MAformer successfully "attended" to periodicities across the entire sequence, capturing the subtidal drift characteristic of the point B in the Outer Shannon Estuary. Table 3 Multi-horizon SSS forecast with CNN-LSTM-GRU and MAformer. Operational Regime Forecast Horizon (h) Model R 2 RMSE MAE Performance Position Short-term 24 Hours CNN-LSTM-GRU 0.892024328 0.149914906 0.108157763 1st MAformer 0.750893488 0.220369513 0.168943027 2nd Medium-term 72 Hours CNN-LSTM-GRU 0.624070932 0.279050657 0.236711443 1st MAformer 0.611834607 0.290250044 0.235846802 2nd 120 Hours CNN-LSTM-GRU 0.518487273 0.333236290 0.276683690 1st MAformer 0.485628361 0.428809332 0.361344060 2nd Long-term 168 Hours CNN-LSTM-GRU 0.319393752 0.364305804 0.299971077 1st MAformer 0.286716687 0.382873008 0.335053076 2nd 240 Hours CNN-LSTM-GRU 0.205662057 0.438079621 0.375534803 2nd MAformer 0.241104769 0.405137999 0.339359355 1st 4.2 Scale-Dependent Ablation and Predictor Dynamics (Hybrid Vs Maformer) The results of the thematic ablation study (Tables 5 – 9 , and the summary in Table 10 ) reveals that the “information value” of environmental drivers is not static; it undergoes a fundamental transformation as the operational horizon expands. The comparative ablation heatmap (Fig. 4 ) reveals a fundamental transition in predictive logic as the forecast horizon expands from 24 to 240 hours. The colour gradient (dark blue to red) further corroborates this pattern, with deeper blue tones (higher R² ≈ 0.75–0.85) concentrated in the Hybrid panels at shorter horizons, indicating stronger explained variance and forecast skill. Error growth with increasing lead time is evident in both models, reflected by the transition toward yellow–orange and red tones (R² < 0.35) at extended horizons. The allied “Regime Shift” is characterized by a transition from Hydrodynamic Dominance to Climatological Anchor Dependence. 4.2.1 Short-Term Regime (24 Hours): Hydrodynamic Dominance In the short-term regime, both architectures demonstrated high resilience to predictor loss, maintaining R 2 values above 0.85. In this regime, the models function as “high-fidelity tidal emulators,” where the salinity flux is a direct product of immediate advection. However, the CNN-LSTM-GRU model performed better than the Maformer in response to all the 5 ablation groups. This further validates the relative superiority of the former in such short-term SSS forecasting in the outer estuary. The Hybrid Table 4 24-hour ablation test between CNN-LSTM-GRU and Maformer. Ablation Group Model R 2 RMSE MAE Model’s Performance (R 2 ) Position Baseline (Full) CNN-LSTM-GRU 0.896933257 0.135205507 0.098653157 1st MAformer 0.851888285 0.212425224 0.163956961 2nd No Steric/Mass CNN-LSTM-GRU 0.885161176 0.150311869 0.100331658 1st MAformer 0.840144642 0.167355447 0.121138977 2nd No Thermal CNN-LSTM-GRU 0.900581700 0.133351443 0.091901315 1st MAformer 0.859673946 0.180258449 0.138860242 2nd No Surface CNN-LSTM-GRU 0.878385488 0.175603426 0.145410965 1st MAformer 0.795323741 0.222047422 0.169220071 2nd No Local Velocity CNN-LSTM-GRU 0.879335760 0.162016699 0.129375088 1st MAformer 0.867726632 0.156752105 0.117856269 2nd model achieved its peak performance (R 2 = 0.901) when the Thermal group (swpt) was ablated, suggesting that at this scale, sea-water potential temperature acts as “climatological noise” that distracts the model from high-frequency tidal signals. In terms of predictor sensitivity, the removal of Local Velocity and Surface Forcing (Wind/Pressure/ssh) caused significant R 2 drops, confirming that 24-hour SSS variability in the Outer Shannon Estuary is primarily a function of tidal momentum and surface stress (Table 4 ). 4.2.2 The Medium-Term Regime (72–120 Hours): The Synoptic Transition The medium-term regime represents the "stability crossover" point. During this 3-to-5-day window, the influence of local hydrodynamic "memory" begins to fade, and the models must rely on synoptic atmospheric trends. In terms of ablation divergence, at 120h, the Hybrid model began to show increased sensitivity to Surface Forcing, while the MAformer showed a more distributed dependency. In terms of information decay, the first sign of the Hybrid model’s (Baseline) performance decay was observed as the R 2 dropped from 0.669 (72h) to 0.585 (120h). This indicates that the recurrent gates are struggling to maintain the phase alignment of tidal cycles over multiple days, marking the beginning of the shift away from deterministic local forcing toward broader environmental trends (Tables 6 and 7 ). Table 5 72-hour ablation test between CNN-LSTM-GRU and MAformer. Ablation Group Model R 2 RMSE MAE Model’s Performance (R 2 ) Position Baseline (Full) CNN-LSTM-GRU 0.668731170 0.266150099 0.216379273 1st MAformer 0.618443024 0.338101862 0.262775844 2nd No Steric/Mass CNN-LSTM-GRU 0.730026473 0.222844677 0.167225976 1st MAformer 0.646609543 0.251102616 0.187637156 2nd No Thermal CNN-LSTM-GRU 0.719551413 0.224270387 0.166829226 1st MAformer 0.649256718 0.279136094 0.223182635 2nd No Surface CNN-LSTM-GRU 0.715667740 0.264021929 0.215648666 1st MAformer 0.598636931 0.272652079 0.204377027 2nd No Local Velocity CNN-LSTM-GRU 0.716426165 0.252263786 0.205253971 1st MAformer 0.650110004 0.254955983 0.195328534 2nd Table 6 120-hour ablation test between CNN-LSTM-GRU and MAformer. Ablation Group Model R 2 RMSE MAE Model’s Performance (R 2 ) Position Baseline (Full) CNN-LSTM-GRU 0.585352424 0.271818918 0.223405082 1st MAformer 0.365146159 0.369032908 0.308609836 2nd No Steric/Mass CNN-LSTM-GRU 0.624782615 0.274110433 0.211652265 1st MAformer 0.399055186 0.361192537 0.287648601 2nd No Thermal CNN-LSTM-GRU 0.397598025 0.327603148 0.239133078 1st MAformer 0.324941169 0.387505403 0.309900979 2nd No Surface CNN-LSTM-GRU 0.620563237 0.274679190 0.234284501 1st MAformer 0.375974473 0.328057605 0.252043118 2nd No Local Velocity CNN-LSTM-GRU 0.583174585 0.266785091 0.209785050 1st MAformer 0.369918966 0.344012224 0.269024928 2nd 4.2.3 Long-Term Regime (168–240 Hours): Climatological Anchor Dependency In the long-term regime, the predictive logic undergoes a total transformation. The "memory" of Local Velocity and short-term Wind stress becomes statistically irrelevant, replaced entirely by oceanic boundary conditions. At the inception of the regime (168 h), the Hybrid model shows relative superiority across all the ablation groups except the Baseline. The regime shift becomes most evident at the 240-hour horizon. Without the Steric/Mass predictors (gmsv, gmmvv), both models show total predictive collapse. Similarly, without the Thermal group predictor (swpt), both models collapsed. However, the No Steric/Mass MAformer’s R 2 dropping from 0.172 to a negligible 0.003; and the No Thermal MAformer’s R 2 collapsing from 0.102 to 0.007 imply a relatively significant anchor effect (Tables 8 and 9 ). This proves that at a 10-day horizon, the Outer Shannon’s salinity is no longer an estuarine problem but a shelf-sea problem. In terms of architectural de-coherence, this regime highlights the MAformer’s strength. Such de-coherence refers to the point where the internal logic of a deep learning model—specifically the CNN-LSTM-GRU—mathematically "unravels" because it can no longer reconcile the physical relationships between the input predictors and the target salinity at a 240-hour lead time. A de-coherence state could be defined through 3 relevant lenses, the Loss of Temporal Continuity (Recurrent Decay); Stochastic Interference (Noise Overpowering Signal); and the "Mean-Reversion" Collapse. While the Hybrid model entered a state of architectural de-coherence—evidenced by the asterisked (*) relatively high R 2 values with relatively high RMSE and MAE values in the Baseline and No Surface groups (Table 8 ); and the asterisked (**) negative importance values (Table 11 )—the MAformer’s attention mechanism successfully isolated the "Climatological Anchors". By "attending" to the global mass signals over the 241-hour window, the MAformer maintained a coherent (though lower) R 2 with relatively low RMSE and MAE values, which imply relatively accurate 240h SSS forecasts (Table 9 ). Table 7 168-hour ablation test between CNN-LSTM-GRU and MAformer. Ablation Group Model R 2 RMSE MAE Model’s Performance (R 2 ) Position Baseline (Full) CNN-LSTM-GRU 0.205159791 0.414166785 0.369023338 2nd MAformer 0.266829291 0.409161521 0.352466184 1st No Steric/Mass CNN-LSTM-GRU 0.524847594 0.376172051 0.313591343 1st MAformer 0.171774763 0.415642084 0.332898032 2nd No Thermal CNN-LSTM-GRU 0.208633724 0.417473004 0.321307549 1st MAformer 0.102111244 0.491008238 0.408948505 2nd No Surface CNN-LSTM-GRU 0.515116858 0.323842764 0.282772665 1st MAformer 0.197248796 0.377983061 0.297896763 2nd No Local Velocity CNN-LSTM-GRU 0.470771902 0.340053825 0.280645513 1st MAformer 0.172934779 0.404019600 0.331836198 2nd Table 8 240-hour ablation test between CNN-LSTM-GRU and MAformer. Ablation Group Model R 2 RMSE MAE Model’s Performance (R 2 ) Position Baseline (Full) CNN-LSTM-GRU 0.219803355* 0.518995208 0.438749539 1st* MAformer 0.215839103 0.447289435 0.352113979 2nd No Steric/Mass CNN-LSTM-GRU 0.023919637 0.473549348 0.388768872 1st MAformer 0.002818044 0.433527423 0.377396571 2nd No Thermal CNN-LSTM-GRU 0.110696059 0.476712087 0.382465886 1st MAformer 0.006665896 0.525376867 0.446542847 2nd No Surface CNN-LSTM-GRU 0.419014486* 0.53503734 0.454317608 1st* MAformer 0.136718881 0.40776054 0.316736389 2nd No Local Velocity CNN-LSTM-GRU 0.35489267 0.402517772 0.333761702 1st MAformer 0.078620458 0.421498015 0.35005752 2nd Table 9 Performance gap between CNN-LSTM-GRU and MAformer in the 240-Hour Ablation. Horizon: 240h Metric Hybrid (Baseline) MAformer (Baseline) Performance Gap Correlation R 2 0.220 0.216 Negligible Error MAE 0.439 0.352 MAformer is 19.8% better Error RMSE 0.519 0.447 MAformer is 13.8% better 4.2.4 Synopsis of the Statistical Behaviors of the Models by Regime I. Short-term Regime (24h) In the 24-hour window, given that the variables like local velocity, sea surface height, and even wind stress are all physically coupled to this tidal pulse, the two models experience Signal Redundancy. If you remove one group (e.g., Velocity), the model can still "infer" the tidal stage from the others. The Hybrid model excels here because its CNN layers are highly efficient at extracting these sharp, repetitive spatial-temporal features from the redundant data stream, leading to its peak performance with R 2 of 0.892–0.901. The MAformer also benefits, but its Attention heads are designed to look for "relationships" rather than just "patterns." In a redundant environment, the MAformer spends computational energy "attending" to multiple variables that are essentially telling it the same thing. This is why the Hybrid slightly outperforms the MAformer at the 24h mark—the Hybrid is a more efficient tool for "simple" redundant signals (Table 10 ). II. Medium-term Regime (72–120h) As we move past 3 days, the "memory" of the initial tidal state begins to blur due to cumulative meteorological forcing (wind-driven mixing). This is the Transition / Decay phase. Statistically, we see the R 2 begin to drop (decay) because the relationship between the input and the target is no longer strictly deterministic. The Hybrid model begins to lose its edge as the LSTM/GRU units struggle with the accumulating phase-lag of the tide, while the MAformer begins to show its strength in identifying the longer-period synoptic "weather" patterns in the data (Table 10 ). III. Long-term Regime (168–240h) At the 10-day mark, the "Redundancy" has evaporated because the local tides and wind gusts are essentially "noise". The wind and velocity no longer correlate with the 10-day salinity trend. The only remaining predictive "signal" is the slow-moving, low-frequency ocean state (Steric Mass/Temperature). This is why the models transition from a state of Signal Redundancy (where they have many ways to win) to Anchor Dependency (where they have only one way to win: the Steric/Thermal data). As the models become Climatological Anchor Dependent, if you ablate these specific variables, the R 2 collapses to near zero as shown in the results (Table 8 ). In effect, the MAformer is the better architecture here because its Global Attention mechanism can bypass the "noise" of 240 hours of tides to focus exclusively on these "anchors," maintaining a stable (albeit lower) absolute error (MAE) (Table 10 ). Table 10 Summary of the operational logic for the Hybrid and MAformer. Regime Primary Driver Better Architecture Statistical Behavior Short-term Local Velocity / Tide Hybrid Signal Redundancy Medium-term Surface Wind / Pressure Hybrid/MAformer Transition / Decay Long-term Steric Mass / Thermal MAformer Climatological Anchor Dependent 4.3 Percentage Importance of Predictor Group for the Regime Shift To further delineate the transition from hydrodynamic to climatological regimes, a Percentage Importance analysis was conducted (Fig. 5 ; Table 11 ). At the 24-hour horizon, the importance of all groups remained below 3%, indicating that the Hybrid model is highly robust to individual sensor failures in the short term. Remarkably, the removal of Thermal data at 24h resulted in a negative importance (-0.4%), confirming that thermal signals act as high-frequency noise for daily tidal predictions. However, at the 240-hour horizon, the Steric/Mass group accounted for 98.6% of the MAformer’s predictive integrity. This identifies a 'Single Point of Failure' for long-term forecasting. While short-term models are resilient and distribute importance across all 9 predictors, the 10-day forecast is entirely reliant on the global mass-change signal. This quantitative shift justifies the operational recommendation of prioritizing global products, particularly including the GOPAF that was leveraged for this study; and relevant satellite-derived products for strategic estuarine planning. Table 11 Percentage importance of predictor groups across operational horizons. Operational Regime Horizon Model Steric/Mass (%) Thermal (%) Surface (%) Local Velocity (%) Short-term 24h Hybrid 1.3% -0.4% 2.1% 2.0% MAformer 1.4% -0.9% 6.7% -1.8% Medium-term 120h Hybrid 32.0% 31.9% -5.8% 0.3% MAformer 25.4% 10.9% -3.0% -1.3% Long-term 240h Hybrid 89.1% 49.5% -90.5% ** -61.4% ** MAformer 98.6% 96.7% 36.6% 63.4% In Table 11 , the asterisked (**) negative values (e.g., -90.5% and − 61.4%) for the Hybrid architecture at the 240h horizon denote a state of architectural de-coherence. At this extreme lead time, the Hybrid model’s recurrent units lose the ability to map long-range dependencies, causing it to overfit to high-frequency residuals in the Surface and Local Velocity groups. The negative importance suggests that the presence of these variables at long horizons introduces 'stochastic interference' for the CNN-LSTM-GRU. This contrasts sharply with the MAformer at the 240h horizon, which maintains positive importance across all groups, demonstrating its superior ability to disentangle noise from signal in the Outer Shannon’s complex multi-regime environment. 4.4 Operational Synthesis: The "Information Cross-Over" By synthesizing both studies, we identify an Information Cross-Over Point. Between 120h and 168h, the models transition from being "Hydrodynamically Driven" to "Climatologically Dependent" (Table 11 ). This explains why the MAformer, which is designed to identify long-range patterns in complex sequences, begins to take the lead in accuracy as the horizon expands to the extreme of 240h. 4.5 Analysis of Metric Divergence: R 2 Stability vs. Error Magnitude A notable finding in the 240-hour regime (Table 8 ) is the divergence between the coefficient of determination (R 2 ) and the absolute error metrics (RMSE and MAE). In the Baseline (Full) group, both architectures exhibited nearly identical R 2 values (Hybrid: 0.219; MAformer: 0.215), yet their error magnitudes differed substantially. The MAformer achieved a 19.64% lower MAE (0.352 vs. 0.438) and a 13.71% lower RMSE (0.447 vs. 0.518) compared to the Hybrid model. This divergence provides critical insight into how each model handles the increased variance of a 10-day horizon in the Outer Shannon Estuary as subsequently detailed. 4.5.1 The "Mean-Reversion" of Hybrid Models The Hybrid model’s higher error magnitude (RMSE, and MAE) despite its relatively high R 2 suggests that the CNN-LSTM-GRU architecture tends to make larger "peak-to-peak" errors. When recurrent units (LSTM/GRU) lose temporal coherence at long horizons, they often revert toward predicting the mean of the training distribution to minimize the global loss function. While this maintains a baseline R 2 (as the model still captures the general trend), it fails to resolve the specific magnitudes of salinity surges or drops, leading to inflated RMSE and MAE. 4.5.2 The "Precision Mapping" of Attention Mechanisms In contrast, the MAformer’s lower error metrics indicate a superior ability to resolve the amplitude of salinity fluctuations. Because the Self-Attention mechanism computes dependencies across the entire 241-hour lookback window without the decay associated with recurrent gates, it can identify specific historical "signatures" (e.g., a high-spring tide combined with a specific wind stress pattern) that correlate with the future state. This results in a forecast that is more closely aligned with the actual observed values on a point-by-point basis, even when the overall explained variance (R 2 ) is constrained by the inherent chaotic nature of the 10-day horizon. 4.5.3 Physical Implication for the Outer Shannon Estuary In an operational context within the Outer Shannon, this divergence is highly significant. An R 2 of 0.21 tells us that both models are capturing the underlying "low-frequency" signal (the Spring-Neap drift). However, the MAformer’s significantly lower MAE implies that it is more reliable for determining whether salinity will cross a specific threshold (critical for aquaculture or desalination) at the 10-day mark. The Hybrid model, while identifying the trend, introduces about 20% more uncertainty in the actual magnitude, which could lead to "false alarms" or missed environmental hazards. 5. Conclusions The study has successfully demonstrated the import of a rigorous comparative study of a Hybrid CNN-LSTM-GRU architecture and a MAformer for an extended multi-horizon (24-240h) SSS forecasting in a relatively complex outer part of Shannon, an European estuary. The outputs of the endurance test offer useful operational information that could support decisions of practitioners in their choice between the two models. The detailed results of the systematic ablation across the 5 horizons help to decode the "information value" of thermal, steric, surface, and velocity predictors at each horizon. This justifies the need for practitioners to ascertain the scale-dependent importance of environmental predictors in the real-world applications of such DL models. The study also addressed the gap of "Information Starvation" by scaling the lookback window to capture up to 19.4 tidal cycles for long-term horizons, a methodology not yet standardized in European estuarine DL research. The result of the endurance test shows the superiority of the Hybrid model for 24–168 hours forecasts; and the superior resilience of MAformer model for 240 hours forecasts. While both models succumb to the "Climatological Collapse" when the anchors are removed, the MAformer demonstrates superior resilience in the Baseline and No Surface groups at 240h. The attention mechanism effectively "filters" the 241-hour lookback window to identify long-period subtidal oscillations. Conversely, the Hybrid model suffers from recurrent information decay, where the high-frequency tidal "noise" from early in the sequence dilutes the long-term trend signal needed for the 10-day forecast. Overall, the results indicate that the Hybrid model provides superior short- and medium-range stability and greater resilience to input perturbation, while MAformer retains competitive long-range capability under complete predictor configurations but is more sensitive to thematic ablation. Thus, this study has successfully established an operational hierarchy for sensor-efficient estuarine monitoring based on forecast lead times. The relevant thematic findings are synthesized as follows. Architectural Trade-offs: The CNN-LSTM-GRU Hybrid is the optimal choice for high-precision, short-term (24h) and medium-term (72-120h) operational alerts. However, the MAformer is more robust for strategic, long-term (240h) forecasting. Regime Sensitivity: Short-term forecasts are inhibited by the "noise" of global thermal/steric variables, whereas long-term forecasts are entirely dependent on them. Windowing Efficacy: The L = h + 1 scaling strategy proved essential, allowing the MAformer to resolve the nearly 20 M2 cycles required to stabilize a 10-day forecast. The Shannon Factor: In the outer Shannon Estuary, the transition from local tidal forcing to open-ocean steric influence occurs between the 3-day and 5-day marks. 6. Recommendations and Future Work 6.1 Operational Recommendations I. Dual-Model Deployment: It is recommended that maritime authorities in the Shannon region adopt a "Hierarchical Forecasting System": a Hybrid model for both daily 24h navigation safety and 72-120h early warning information/preparedness; and an MAformer for 10-day environmental management. II. Sensor Prioritization: Investment should be prioritized for high-frequency velocity sensors for daily operations, but real-time integration with global sea-level mass data (Steric/Mass) is mandatory for any long-term planning tools. 6.2 Future Study Proposals I. Extreme Event Analysis: Future work should investigate how these architectures perform during extreme storm surges or "1-in-50-year" fluvial flood events in the Shannon. II. Transfer Learning: A study should be conducted to see if the "Climatological Anchor" discovered here for the Shannon can be transferred to other European estuaries (e.g., the Elbe or Tagus) to reduce training time. 6.3 Study Limitations: Operational Boundaries and Architectural De-coherence While this study establishes the MAformer as a resilient tool for long-term forecasting, it is essential to define the boundaries within which these findings remain valid. Every data-driven model possesses a "prediction horizon limit" beyond which physical-mathematical alignment fails. I. The De-coherence Threshold As identified in subsections 4.2.3 and 4.3, the CNN-LSTM-GRU Hybrid reaches a state of architectural de-coherence at the 240-hour mark. This represents a fundamental limitation: recurrent architectures are constrained by their "memory bottleneck." Researchers should be cautioned that extending such models beyond the 10-day limit in macrotidal environments like the Outer Shannon Estuary may result in "stochastic interference," where the model produces outputs based on residual noise rather than actual hydrodynamic drivers. II. Dependency on the "Climatological Anchor" The primary limitation of the MAformer—and indeed any model operating at the 240h time scale—is its absolute dependency on the Steric/Mass (gmsv) and Thermal (swpt) groups. As the ablation study proved, these variables act as the "Climatological Anchor". However, the constraint is that if satellite-derived steric data or deep-water temperature sensors experience a telemetry failure, the model's predictive power for horizons >120h will collapse entirely. The model does not have a "physics-backup" to compensate for the loss of these specific global signals. III. Spatial Specificity of the "Outer" Estuary This study specifically targeted the Outer Shannon, where salinity is dominated by open-ocean exchange and shelf-sea interactions; particularly at the point where the AFWA is relatively high. However, there is a constraint. These findings may not be directly transferable to the Inner Shannon (river-dominated) without recalibration. In the inner estuary, river discharge (fluvial forcing) likely replaces Steric/Mass as the primary "anchor" for long-term horizons. Therefore, the architectural success of the MAformer here is tied to its ability to process oceanic signals, and its performance under river-dominant regimes remains an area for future validation. IV. Data Density and Frequency The L=h+1 windowing strategy requires high-density, uninterrupted time-series data. In many operational settings, "data gaps" are common. The current architectures are sensitive to these gaps; the MAformer, in particular, requires a complete 241-hour sequence to calculate its Attention Heads effectively. The limitation lies in the "data-hungry" nature of Transformer-based models compared to simpler statistical regressions. Declarations Author Contributions: Conceptualization, O.A-J.; methodology, O.A-J.; software, O.A-J.; validation, O.A-J.; formal analysis, O.A-J.; investigation, O.A-J.; resources, O.A-J.; data curation, O.A-J. (data downloaded from EU CMEMS repository); writing—original draft preparation, O.A-J.; writing—review and editing, O.A-J. and O.A-J.; visualization, O.A-J.; supervision, O.A-J.; project administration, O.A-J. The author has read and agreed to the published version of the manuscript. Funding: This research received no external funding. Informed Consent Statement: Not applicable. Data Availability Statement: The dataset used was obtained from the EU CMEMS via https://data.marine.copernicus.eu/product/GLOBAL_ANALYSISFORECAST_PHY_001_024/download?dataset=cmems_mod_glo_phy_anfc_0.083deg_PT1H-m_202406 (Data); https://data.marine.copernicus.eu/product/GLOBAL_ANALYSISFORECAST_PHY_001_024/download?dataset=cmems_mod_glo_phy_anfc_merged-sl_PT1H-i_202411 (Data); https://data.marine.copernicus.eu/product/GLOBAL_ANALYSISFORECAST_PHY_001_024/download?dataset=cmems_mod_glo_phy_anfc_merged-uv_PT1H-i_202211 (Data); and https://data.marine.copernicus.eu/product/GLOBAL_ANALYSISFORECAST_PHY_001_024/description or https://doi.org/10.48670/moi-00016 (Metadata documentation). Conflicts of Interest: The authors declare no conflict of interest. Clinical Trial Number: Not applicable. Ethics and Consent to Participate Declarations: Not applicable. References Ajibola-James, O. (2026). Investigation of High-Frequency Dynamics of Sea Surface Salinity in the Outer Shannon Estuary Using Numerical Model-Derived Data. Preprints . https://doi.org/10.20944/preprints202601.1029.v1 Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 . Costello, M. J. (2025). Ecological responses to salinity variability in Irish estuarine systems. Estuarine, Coastal and Shelf Science, 312 , 108945. Gorski, G., Cook, S. E., Snyder, A. M., Appling, A. P., Thompson, T. P., Smith, J. D., Warner, J. C., & Topp, S. N. (2024). Deep learning of estuary salinity dynamics is physically accurate at a fraction of hydrodynamic model computational cost. Limnology and Oceanography , 69(5), 1070–1085. https://doi.org/10.1002/lno.12549 Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation , 9(8), 1735-1780. Nguyen, H., et al. (2025). Estuary salinity prediction using machine learning: Case study in the Hau Estuary in Mekong River. Water Supply . https://doi.org/10.2166/ws.2025.007 Ning, P., Zhang, C., Zhang, X., & Jiang, X. (2021). Short- to medium-term sea surface height and salinity prediction using an optimized simple recurrent unit deep network. Frontiers in Marine Science , 8 , 672280. O'Donncha, F., Grant, J., & Hill, R. J. (2020). An environmental monitoring and forecasting system for sustainable marine aquaculture. Journal of Marine Systems , 201, 103239. https://doi.org/10.1016/j.jmarsys.2019.103239 Pritchard, D. W. (1967). What is an estuary: physical viewpoint. Estuaries , 83, 3-5. Qi, S., He, M., Hoang, R., Zhou, Y., Namadi, P., Tom, B., Sandhu, P., Bai, Z., Chung, F., & Ding, Z. (2023). Salinity modeling using deep learning with data augmentation and transfer learning. Water , 15(13), 2482. https://doi.org/10.3390/w15132482 Saccotelli, G., et al. (2024). Enhancing estuary salinity prediction: A machine learning and deep learning based approach. Applied Computing and Geosciences , 23, 100173. https://doi.org/10.1016/j.acags.2024.100173 Schmidt, A., Ziegeler, M., & Mayer, B. (2021). Hybrid CNN-LSTM networks for Sea Surface Salinity in the German Bight. Ocean Modelling , 162, 101804. Zhang, Y., et al. (2025). Predictions of saltwater intrusion in the Changjiang Estuary: Integrating machine learning methods with FVCOM. Journal of Hydrology , 653, 132739. https://doi.org/10.1016/j.jhydrol.2025.132739 Zheng, R., Sun, Z., Jiao, J., Ma, Q., & Zhao, L. (2024). Salinity prediction based on improved LSTM model in the Qiantang Estuary. Journal of Marine Science and Engineering , 12(8), 1339. https://doi.org/10.3390/jmse12081339 Zhu, B., Wang, T., De Meester, J., & Willems, P. (2024). Comparative analysis with statistical and machine learning for modeling salinity along the Scheldt Estuary. Water , 16(15), 2150. https://doi.org/10.3390/w16152150 Additional Declarations The authors declare no competing interests. Supplementary Files Appendixs.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9178680","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":609456260,"identity":"551d45fd-e8b6-4c8c-8823-6733dd4900e7","order_by":0,"name":"Opeyemi Ajibola-James","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0002-3012-7569","institution":"Geo Inheritance Ltd","correspondingAuthor":true,"prefix":"","firstName":"Opeyemi","middleName":"","lastName":"Ajibola-James","suffix":""}],"badges":[],"createdAt":"2026-03-20 11:57:33","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9178680/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9178680/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105209711,"identity":"e0d83858-fb83-4adf-8d97-2d90d40b4ed3","added_by":"auto","created_at":"2026-03-23 13:28:13","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":521690,"visible":true,"origin":"","legend":"\u003cp\u003eMap of the study area showing the monitoring point B in relation to points A, C, and D.\u003c/p\u003e\n\u003cp\u003eSource: Ajibola-James (2026).\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9178680/v1/890755442d4fd695d01e294b.png"},{"id":105752085,"identity":"4a591b7d-aecb-40ec-9d43-6f1a3148e20f","added_by":"auto","created_at":"2026-03-30 15:54:37","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":42026,"visible":true,"origin":"","legend":"\u003cp\u003eVariable Inflation Factor.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9178680/v1/c693fef4c72ca745c128d711.png"},{"id":105209800,"identity":"8dd8d640-1642-4164-ae28-a72e34d196ba","added_by":"auto","created_at":"2026-03-23 13:28:33","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":143457,"visible":true,"origin":"","legend":"\u003cp\u003eThematic ablation workflow diagram.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9178680/v1/49c4c5e32f1b89eeab8607ec.png"},{"id":105209733,"identity":"e24a3579-9a94-44fd-89aa-2f65f8f1a931","added_by":"auto","created_at":"2026-03-23 13:28:21","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":226686,"visible":true,"origin":"","legend":"\u003cp\u003eComparative ablation heatmap showing the scale-dependent regime shift.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9178680/v1/6d47e049514bc120c89bc88c.png"},{"id":105209707,"identity":"89160096-7c95-4f60-81c6-79365c12d052","added_by":"auto","created_at":"2026-03-23 13:28:11","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":63121,"visible":true,"origin":"","legend":"\u003cp\u003ePercentage importance of predictor group for the regime shift from hydrodynamic dominance to climatological anchor dependence.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9178680/v1/3f0e8292c184fa6b074086cf.png"},{"id":105752627,"identity":"98ac3e95-efaf-4ab2-9024-225da8d2b8fa","added_by":"auto","created_at":"2026-03-30 16:03:13","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3085893,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9178680/v1/eba92eb1-ed8b-4c30-826e-1db326b2480b.pdf"},{"id":105209704,"identity":"bb54630b-b055-42d3-98be-10e5d0ec00d0","added_by":"auto","created_at":"2026-03-23 13:28:11","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":290085,"visible":true,"origin":"","legend":"","description":"","filename":"Appendixs.docx","url":"https://assets-eu.researchsquare.com/files/rs-9178680/v1/ecfcfdf6ec8ed3aa9a260bf8.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eA Comparative Ablation Study of CNN-LSTM-GRU and MAformer Architectures for Operational Multi-regime Salinity Forecasting in the Outer Shannon Estuary\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eEstuaries represent some of the most complex and dynamic aquatic environments on Earth. As transitional zones between fluvial freshwater systems and the saline open ocean, they are characterized by intense gradients in salinity, temperature, and nutrient concentration (Pritchard, 1967). SSS is perhaps the most critical parameter in these systems, serving as a primary indicator of water density, circulation patterns, and habitat suitability for diverse marine species. Accurate SSS forecasting is essential for operational decisions in desalination, maritime navigation, and the management of sensitive aquaculture sites. Historically, estuarine modelling has relied on numerical hydrodynamic models, such as the Finite-Volume Community Ocean Model (FVCOM) or the Regional Ocean Modelling System (ROMS). While these models are physically consistent, they are computationally expensive and often struggle with the \"curse of dimensionality\" when incorporating high-frequency local sensor data alongside global satellite products. Recent advancements in computational hydrology have shifted toward the application of deep learning (DL) architectures to capture the non-linear and multi-scale temporal dependencies inherent in estuarine salinity dynamics. Models like Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRU) have demonstrated an exceptional ability to capture the temporal dependencies of tidal cycles (Hochreiter \u0026amp; Schmidhuber, 1997; Cho et al., 2014). The LSTM networks have emerged as a foundational tool due to their ability to mitigate the vanishing gradient problem in long-term environmental time series. Gorski et al. (2024) demonstrated that LSTM-based models could achieve the physical accuracy of traditional 3D hydrodynamic models in the Chesapeake and Delaware Bays while operating at a fraction of the computational cost. Similarly, Saccotelli et al. (2024) validated the efficacy of LSTMs in predicting salt wedge intrusions in the Po River Delta, highlighting their superior performance in micro-tidal environments compared to traditional machine learning algorithms.\u003c/p\u003e\n\u003cp\u003eTo further enhance predictive stability, researchers have begun integrating hybrid DL structures and meta-heuristic optimization algorithms. Qi et al. (2023) expanded this by introducing residual connections (Res-LSTM and Res-GRU) combined with transfer learning to address data scarcity and improve generalizability across different estuarine regimes. Nguyen et al. (2025) proposed a hybrid LSTM-GRU model optimized with the Sailfish Optimizer (SFO) for the Hau Estuary, reporting significant improvements in 7-day forecast horizons. The integration of GRU is often preferred for its reduced parameter complexity without sacrificing the performance of gated mechanisms. The trend toward \"Physics-Informed\" or \"Hybrid-Mechanism\" modelling is also gaining traction. Optimization of hyperparameters remains a critical focal point; Zheng et al. (2024) utilized an Improved Sparrow Search Algorithm (ISSA) to tune LSTM parameters in the Qiantang Estuary, yielding about 10% reduction in RMSE. Complementary to these temporal models, Zhu et al. (2024) emphasized the role of data preprocessing, showing that Discrete Wavelet Transform (DWT) decomposition within an Artificial Neural Network (ANN) framework significantly sharpens the model’s sensitivity to extreme high-salinity events triggered by sea-level rise. Zhang et al. (2025) coupled DL models with the FVCOM in the Changjiang Estuary, finding that the DL component reduced numerical simulation errors by over 50% by learning the residual biases of the physical model.\u003c/p\u003e\n\u003cp\u003eDespite these advances, some significant research gaps remain. First, the European coastline and the allied outer estuaries that are characterized by complex tidal regimes and diverse freshwater discharge patterns (particularly including the outer parts of Shannon, and Tagus estuaries) are still understudied with tripartite variants of such hybrid data-driven models, particularly CNN-LSTM-GRU; and evolving generalist models for time series forecasting, particularly MAformer. Second, most DL salinity models still focus on short-term horizons (e.g., 24-48 hours), and a few focus on medium-term horizons (e.g., 72–120 hours). There is a dearth of research into how the predictive power of DL architectures decays over \"Endurance Tests\" extending to 10 days (240 hours), a timeframe critical for long-term water resource planning. Third, the relative importance of environmental predictors is often assumed to be static. However, estuarine physics suggests that the drivers of a 24-hour forecast (dominated by M2 tides) differ fundamentally from the drivers of a 240-hour forecast (influenced by Spring-Neap transitions and global steric sea-level changes). This suggests lack of scale-dependent ablation. There is no rigorous evidence in current literature that demonstrates how the \"information value\" of a specific predictor (e.g., Steric vs. Thermal) shifts as the forecast moves from a tidal regime (24h) to a climatological regime (240h).\u0026nbsp;The gaps present apparent need for a rigorous comparative study of a Hybrid CNN-LSTM-GRU architecture and a MAformer for an extended multi-horizon (24-240 hours) SSS forecasting in a relatively complex outer part of European estuary. This should involve a systematic ablation study across five horizons that could decode the \"information value\" of thermal, steric, surface, and velocity predictors. Additionally, such study should address the gap of \"Information Starvation\" by scaling the lookback window to capture up to 19.4 tidal cycles for long-term horizons, a methodology not yet standardized in European estuarine DL research. In this regard, the objectives of this study are to: (i) evaluate the performance decay of Hybrid and MAformer architectures across 1-day (24-h) to 10-day (240-h) forecast horizons; (ii) identify the scale-dependent importance of environmental predictors through thematic ablation; and (iii) establish an operational hierarchy for sensor-efficient estuarine monitoring based on forecast lead times.\u003c/p\u003e"},{"header":"2. Study Area","content":"\u003cp\u003eThe Shannon Estuary, the largest estuarine system in Ireland, extends approximately 100 km from Limerick City westward to the Atlantic Ocean between Loop Head (County Clare) and Kerry Head (County Kerry), encompassing a navigable surface area of roughly 500 km². It forms the tidal outlet of the River Shannon—Ireland’s longest river—and represents a classic macrotidal environment characterized by strong tidal forcing, complex bathymetry, and pronounced salinity gradients. The estuary exhibits highly dynamic hydrographic conditions, where rapid fluctuations in sea surface salinity (SSS) are a defining physical feature and an important ecological control (Costello, 2025). The monitoring (observations) point used in this study is situated in the outer estuary (Figure 1), where marine–fluvial interactions are particularly intense.\u0026nbsp;The outer part of the Shannon Estuary is characterised by significant spatial and temporal dynamics, which underscore a complex hydrographic regime characterised by distinct geographic gradients and pronounced seasonal transitions\u0026nbsp;(Ajibola-James, 2026).\u003c/p\u003e\n\u003cp\u003ePoints A, B, C, and D have annual mean SSS (AMSSS) of 33.985, 33.881, 34.125, and 34.343; and annual mean CV (AMCV) of 0.086, 0.073, 0.094, and 0.106 % respectively (Figure 1). The lowest values exhibited by the point B\u0026nbsp;(52.500000\u003csup\u003eo\u003c/sup\u003e N, -9.74999\u003csup\u003eo\u003c/sup\u003e W), the data observations point imply the highest level of annual freshwater availability (AFWA); and the most stable SSS on annual time scale (Ajibola-James, 2026). Salinity variability within the Shannon Estuary plays a critical role in structuring biological processes and ecosystem functioning. In particular, marked SSS oscillations in the inner estuary have been shown to influence lipid and fatty acid composition in key estuarine biota, highlighting salinity as a primary environmental stressor (Costello, 2025). Beyond its ecological relevance, the estuary and its catchment have substantial socio-economic importance, especially for Ireland’s Mid-West Region. The River Shannon supports navigation, fisheries, industry, and potable water supply; however, this resource is vulnerable to upstream seawater intrusion (USI), which poses risks to freshwater abstraction and water security under changing hydrodynamic and climatic conditions.\u003c/p\u003e"},{"header":"3. Methodology","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Datasets and the Observation Point\u003c/h2\u003e \u003cp\u003eThe most appropriate data observation point in the outer Shannon Estuary for this study was statistically determined with the most relevant recent finding on the SSS variability by Ajibola-James (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2026\u003c/span\u003e) that established the point B as the most appropriate for monitoring and predicting USI in the outer estuary. Additionally, the need for making this a foremost task in this SSS prediction study is in consonance with his opinion. The study utilised datasets collected from the point B, incorporating SSS and 11 potential predictors (01 Mar. 2024, 00:00 hr-30 Jun. 2025, 23:00 hr), Global mean sea-level variations (gmsv), Global mean mass volume variation (gmmvv) that is also called Global average sea level change due to water mass; Sea water potential temperature (swpt); Inverted barometer (ib) that depicts Change in sea surface height due to change in air pressure; the allied wind speed components, particularly Sea surface water stokes drift x velocity (vsdx) and Sea surface water stokes drift y velocity (vsdy); sea surface height above geoid (ssh); Northward sea water velocity (nswv); Eastward sea water velocity (eswv); Sea surface height above geoid due to ocean tide (otide); and Surface sea water y velocity due to tide (vtide). The datasets were retrieved from the online data repository of the European Union (EU) Copernicus Marine Service Information (CMEMS) using the Global Ocean Physics Analysis and Forecast (GOPAF) product, L4 (hourly).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Data Cleaning: Data Quality Control and Anomaly Remediation\u003c/h2\u003e \u003cp\u003ePreliminary inspection of the 16-month datasets at hourly time scale revealed that missing values (NAs) were minimal (approx. 3.06%); and 2 predictor variables, specifically the velocity components (vsdx and vsdy) exhibited 'hard zero' anomalies. These were defined as continuous sequences of 0.000 values exceeding two hours\u0026mdash;a state physically inconsistent with the dynamic tidal regime of the Shannon Estuary. The order of the anomaly remediation was characterized by zero-to-NA followed by hybrid imputation. To ensure temporal continuity for the deep learning architectures, a hybrid imputation strategy was implemented using the \u003cem\u003eimputeTS\u003c/em\u003e framework in R software. The \u003cem\u003epacman\u003c/em\u003e, and \u003cem\u003etidyverse\u003c/em\u003e libraries in R were also utilised during the data cleaning tasks. Isolated missing points and short-duration gaps (less than 10 hours) were remediated using linear interpolation. In contrast, larger gaps (above 10, reaching up to 30 hours in the 2 variables) were reconstructed using State-Space Kalman Smoothing (based on a structural time-series model). This approach preserves the underlying sinusoidal tidal harmonics, ensuring that the CNN-LSTM-GRU filters and MAformer attention heads receive a physically coherent signal rather than artificial step-functions or flat-line artifacts. This was achieved prior to the implementation of the three-stage variable selection pipeline with appropriate R libraries to ensure that the features selection was driven by physical dependencies rather than sensor artifacts. Technically, this ensured that the information-theoretic metrics (Mutual Information) and recursive elimination (RFE) operated on a continuous and physically representative state-space.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Multicollinearity Detection\u003c/h2\u003e \u003cp\u003eTo ensure that this study meets the highest standards of statistical rigor and hydrological/oceanographic operations, a Variance Inflation Factor (VIF) analysis with appropriate R libraries (\u003cem\u003epacman, car, tidyverse\u003c/em\u003e) coupled with the author\u0026rsquo;s Hydrologic/Oceanographic domain intelligence were leveraged to detect the inherent multicollinearity in the 11 potential SSS predictors. VIF (Multivariate) measures how much the variance of a regression coefficient is inflated due to collinearity with all other predictors in the model. It detects \"collective\" overlap that a simple correlation matrix (e.g. Pairwise Correlation) might miss. The VIF decision rules usually involve removal of high multicollinearity variables with VIF\u0026thinsp;\u0026gt;\u0026thinsp;10; and retention of the moderate multicollinearity variables (VIF\u0026thinsp;\u0026gt;\u0026thinsp;5 but \u0026lt;\u0026thinsp;10); and relatively low multicollinearity variables (VIF\u0026thinsp;\u0026lt;\u0026thinsp;5). The initial multicollinearity assessment via the VIF indicated that all the 11 potential predictors remained below the critical threshold of 10, with gmsv (7.8767) exhibiting the highest collinearity. Notably, astronomical tidal height (otide) and tidal velocity (vtide) displayed VIF values near 1.01, suggesting statistical independence (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). In the context of the Shannon Estuary, variables like different velocity components or overlapping pressure indices often move so closely together that they can easily mask each other's true importance (Zhang et al., 2021), particular including in VIF analysis. Consequently, the 2 predictors with the relatively moderate VIF, gmsv (7.8767) and gmmvv (4.9714) that could have been erroneously deleted by the researcher to achieve the goal of retaining 9 robust predictors for the deep learning modelling were retained. The decision to retain them till the time of implementing the relevant 3-stage feature selection pipeline was based on the author\u0026rsquo;s Hydrologic/Oceanographic domain intelligence as subsequently detailed in 3.4.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Feature Selection (The 3-stage Pipeline)\u003c/h2\u003e \u003cp\u003eThe utilised 3-stage feature selection pipeline was implemented with relevant R libraries (\u003cem\u003epacman, praznik\u003c/em\u003e, \u003cem\u003ecaret\u003c/em\u003e, \u003cem\u003exgboost\u003c/em\u003e, \u003cem\u003etidyverse\u003c/em\u003e, \u003cem\u003escales\u003c/em\u003e, and \u003cem\u003erandomForest\u003c/em\u003e). The final predictor suite was determined by a consensus ranking of three distinct feature selection paradigms: (a) Mutual Information, representing non-linear statistical dependence; (b) RFE, representing optimal subset performance via 10-fold cross-validation; and (b) XGBoost Gain, representing decision-tree-based predictive contribution. These three different mathematical perspectives are usually depicted by a single defensible truth called Global Score (GS). High GS (e.g., \u0026gt; 0.8) indicates that the variable performed exceptionally well across all three tests. This is the universal importance index, which implies that the variable is a \"Primary Driver\" of salinity. Moderate GS (e.g., 0.4\u0026ndash;0.7) indicates that the variable might be very strong in non-linear relationships (High MI) but slightly less efficient at tree-splitting (Lower XGB). This is the conditional importance index, which implies that the variable is a \"Secondary Driver\". Low GS (e.g., \u0026lt; 0.3) indicates that the variable provides some value but is likely on the verge of being \"noise\" rather than \"signal\". This is the marginal importance index, which implies that the variable is a \u0026ldquo;Minor Driver\u0026rdquo;. As shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the 'Robust 9' predictors (green with Rank 1\u0026ndash;9) demonstrated high convergence across all three metrics. Notably, the 2 variables with the relatively high but moderate multicollinearity, gmsv (VIF\u0026thinsp;=\u0026thinsp;7.8767) and gmmvv (VIF\u0026thinsp;=\u0026thinsp;4.9714) were objectively retained only because their relatively high GS (0.9696, and 0.6594 respectively) ranked within the top 9. Conversely, the otide (VIF\u0026thinsp;=\u0026thinsp;1.0113) and vtide (VIF\u0026thinsp;=\u0026thinsp;1.0082) that showed the least multicollinearity were automatically eliminated due to their relatively low GS rank of 0.0099 and 0.0037 (yellow with Rank 10\u0026ndash;11) respectively. The use of the GS rank as the final assessment index was to ensure that the selected feature space prioritized predictive power while maintaining numerical stability for the MAformer and CNN-LSTM-GRU models.\u003c/p\u003e \u003cp\u003eIt should be underscored that otide and vtide are mathematical constituents (harmonics), not physical \"drivers\" like wind stress or potential temperature. The automatic exclusion of the 2 potential predictors is also in alignment with the \"Thematic Ablation\" Framework that is developed and utilized in this study, which grouped the SSS predictors into Steric/Mass, Thermal, Surface, and Local Velocity. Scientifically and realistically, to maintain the integrity of a Thematic Ablation Study, the predictors must represent distinct physical processes in the real-world. Astronomical tides are a constant background oscillation; the variability in SSS\u0026mdash;which is the target variable\u0026mdash;is driven by the 9 predictors that are objectively retained. The intricate tidal-fluvial interaction within the Shannon makes velocity components particularly prone to mutual dependency, requiring careful feature selection to reveal their distinct roles in salinity advection (O'Donncha et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The exclusion of the 2 predictors focuses the architectures on the primary physical drivers of salinity transport\u0026mdash;Steric, Thermal, Surface, and Velocity regimes\u0026mdash;while preventing the Attention mechanism from over-fitting to theoretical harmonics. Thus, running such a VIF in accordance with appropriate domain knowledge in Hydrology/Physical Oceanography before the subsequent RFE and XGBoost stages helps to clean the field of redundant predictor variables.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRank of the SSS potential predictors based on the Global Score.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRank\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePredictor Variable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMI\u003c/p\u003e \u003cp\u003eScore\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAvgImportance\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRFE\u003c/p\u003e \u003cp\u003eNorm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eXGB\u003c/p\u003e \u003cp\u003eGain\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMI\u003c/p\u003e \u003cp\u003eNorm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eXGB\u003c/p\u003e \u003cp\u003eNorm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eGlobal\u003c/p\u003e \u003cp\u003eScore\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003egmsv\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.624417\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e54.83446\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.908815\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.622876\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.000000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e1.000000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.969605\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003egmmvv\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.435738\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e58.57355\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.000000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.179921\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.689391\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.288662\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.659351\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eswpt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.444438\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e48.20372\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.747112\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.172262\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.703714\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.276363\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.57573\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eib\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.086757\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e43.54378\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.633471\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.010145\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.114891\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.01602\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.254794\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003evsdx\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.086377\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e36.42265\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.459808\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.008185\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.114264\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.012873\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.195649\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003essh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.037979\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e37.95439\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.497163\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.002237\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.034591\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.003321\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.178358\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003evsdy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.032835\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e34.64113\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.416363\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.001991\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.026122\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.002926\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.14847\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enswv\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.019623\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e25.51207\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.193733\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.001169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.004372\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.001607\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.066571\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eeswv\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.017934\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e25.13565\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.184554\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.000788\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.001593\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.000995\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.062381\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eotide\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.016967\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e18.78234\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.029616\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.000258\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.000000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.000144\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.00992\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003evtide\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.023656\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e17.56791\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.000000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.000169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.011012\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.000000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.003671\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Conceptual Framework: Endurance Test Between CNN-LSTM-GRU and MAformer\u003c/h2\u003e \u003cp\u003eThe first part of the modelling study was focused on endurance test between CNN-LSTM-GRU and MAformer, which leveraged the 9 robust predictors for multi-regime SSS forecasting (24, 72, 120, 168, and 240 hours ahead). The appropriate R libraries, \u003cem\u003epacman\u003c/em\u003e, \u003cem\u003etidyverse\u003c/em\u003e, \u003cem\u003ekeras3\u003c/em\u003e, \u003cem\u003etensorflow\u003c/em\u003e, \u003cem\u003eMetrics\u003c/em\u003e, \u003cem\u003eggplot2\u003c/em\u003e, \u003cem\u003escales\u003c/em\u003e, and \u003cem\u003ereshape2\u003c/em\u003e were leveraged to achieve this part. The utilised conceptual framework serves as the \"logical blueprint\" of this part of the research by bridging the gap between environmental physics (the predictors), computational engineering (the architectures), and operational decision-making (the horizons). The framework was built on a \"Multiscale Predictive Logic\" that explains why the model's structure must change as the forecast horizon expands. It is structured into four integrated modules, (a) Environmental Forcing, (b) Temporal Scaling, (c) Architectural Processing, and (d) Operational Evaluation as subsequently detailed. This framework is designed to move beyond simple \"black-box\" forecasting by establishing a physically grounded relationship between input lookback and output lead time.\u003c/p\u003e \u003cp\u003e(a) Environmental Forcing Module (The Inputs)\u003c/p\u003e \u003cp\u003eThe framework begins by categorizing the 9 robust predictors into three physical \"influence layers\" that drive salinity flux in the Outer Shannon Estuary:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eLayer 1 (The Local Pulse): High-frequency hydrodynamic signals consisting of local seawater velocity (nswv, eswv) and total sea surface height (ssh).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eLayer 2 (The Surface Forcing): Synoptic atmospheric variables including wind stress components (vsdx, vsdy) and inverted barometer pressure (ib).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eLayer 3 (The Climatological Anchor): Deep-memory variables including seawater potential temperature (swpt) and global steric/mass variations (gmsv, gmmvv).\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003e(b) Temporal Scaling Module (The Logic)\u003c/p\u003e \u003cp\u003eTo capture the cyclic nature of the estuary, the framework applied a physics-proportional windowing/scaling for the lookback window (L) relative to the forecast horizon (h) such that L\u0026thinsp;=\u0026thinsp;h + 1. For example, for the 24-hour study, L\u0026thinsp;=\u0026thinsp;25 (capturing 2 tidal cycles), while for the 240-hour study, L\u0026thinsp;=\u0026thinsp;241 (capturing\u0026thinsp;~\u0026thinsp;19.4 tidal cycles). This logic is critical for the \"Endurance Test,\" as it provides the architecture with approximately 19.4 M2 tidal cycles, allowing the model to \"see\" the transition between Spring and Neap tidal phases before making a relatively long-term SSS prediction (10 days in the future).\u003c/p\u003e \u003cp\u003e(c) Architectural Processing Module (The Engines)\u003c/p\u003e \u003cp\u003eThe two distinct computational philosophies contrasted in the study are as follows:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eHybrid (CNN-LSTM-GRU): It focuses on Spatio-Temporal Feature Extraction. The CNN layers filter local spatial gradients, while the LSTM and GRU units handle short-to-medium range temporal \"memory\". This architecture consists of a 1D-Convolutional layer (64 filters, kernel size 3) for local pattern recognition, followed by an LSTM layer (100 units) and a GRU layer (50 units) to capture hierarchical temporal features. A dropout rate of 0.2 was applied to mitigate overfitting.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eMAformer (Attention-based): Focuses on Long-Range Dependency Mapping. The Multi-Head Attention mechanism identifies non-linear correlations across the entire 241-hour sequence, bypassing the vanishing gradient limitations of traditional recurrent units. The MAformer employs a Multi-Head Attention mechanism with 4 heads and a key dimension of 25. This allows the model to simultaneously attend to different periodicities (e.g., the 12.42h tidal cycle and the 24h solar cycle). The architecture includes Layer Normalization and Feed-Forward Networks to stabilize the learning process.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003e(d) Operational Evaluation Module (The Regimes)\u003c/p\u003e \u003cp\u003eThe final stage of the framework groups the five horizons in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e into three operational regimes, allowing for a strategic interpretation of the results, short-term (24h) that focused on navigational safety and tidal advection; the medium-term (72h\u0026ndash;120h) that focused on operational dredging and synoptic weather planning; and long-term (168h\u0026ndash;240h) that focused on strategic habitat management and climatological drift.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eOperational regimes and configuration for the multi-regime SSS forecasting.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOperational Regime\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eForecast Horizon (h)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOptimal Lookback (L)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTidal Cycles (approx. M2​)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePhysical Significance \u0026amp; Scaling Logic\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eShort-term\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e~\u0026thinsp;2.0 Cycles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eResolves daily semidiurnal rhythms and immediate tidal advection.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMedium-term\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e72 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e73 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e~\u0026thinsp;5.9 Cycles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCaptures synoptic weather shifts and cumulative tidal phase drift.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e120 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e121 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e~\u0026thinsp;9.7 Cycles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBalances high-frequency data with 5-day estuarine \"memory.\"\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eLong-term\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e168 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e169 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e~\u0026thinsp;13.6 Cycles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCaptures nearly a full 14-day Spring-Neap tidal transition.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e240 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e241 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e~\u0026thinsp;19.4 Cycles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eResolves climatological trends and full Spring-Neap cycles.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Standardized Training Protocol and Error Metrics\u003c/h2\u003e \u003cp\u003eTo ensure a fair comparison across all horizons and ablation groups, all the models utilised in the first and second parts of the modelling study were trained with:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eEpochs: 20\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ePatience: 12 (Early Stopping)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eOptimizer: Adam\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eLoss Function: Mean-Squared-Error (MSE)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eSimilarly, the accuracy of all the models was assessed with the coefficient of determination (R\u003csup\u003e2\u003c/sup\u003e); and the absolute error metrics, root-mean-squared-error (RMSE) and mean-squared-error (MAE). The higher the value of R\u003csup\u003e2\u003c/sup\u003e, the better a model\u0026rsquo;s performance in terms of variation explained; the lower the RMSE and MAE, the better a model\u0026rsquo;s forecasts accuracy.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.7 Conceptual Framework: Ablation Test Between CNN-LSTM-GRU and MAformer\u003c/h2\u003e \u003cp\u003eThe second part of the modelling study was focused on ablation test between CNN-LSTM-GRU and MAformer, which also leveraged the same 9 objectively-selected predictors for the baseline, multi-regime SSS forecasting (24, 72, 120, 168, and 240 hours ahead). The suitable R libraries, \u003cem\u003epacman\u003c/em\u003e, \u003cem\u003etidyverse\u003c/em\u003e, \u003cem\u003ekeras3\u003c/em\u003e, \u003cem\u003etensorflow\u003c/em\u003e, \u003cem\u003eMetrics\u003c/em\u003e, \u003cem\u003eggplot2\u003c/em\u003e, \u003cem\u003escales\u003c/em\u003e, and \u003cem\u003ereshape2\u003c/em\u003e were leveraged to achieve this part. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents the thematic ablation workflow diagram. Generally, a credible ablation study helps to \"unmask\" the predictors importance that simple correlation matrices (e.g. Pairwise Correlation) might miss. The ablation study reveals how predictor importance shifts across the horizons. The 9 robust predictors were categorized into four thematic groups namely (a) Steric/Mass (gmsv, and gmmvv), (b) Thermal (swpt), (c) Surface (ib, vsdx, vsdy, and ssh), and (d) Local Velocity (nswv, and eswv) for the ablation study. In effect, the ablation group for the multi-regime SSS forecasting consists of Baseline (Full), No Steric/Mass, No Thermal, No Surface, and No Local Velocity. To make the results of the ablation test comparable for supporting relevant operational decisions, the same \u0026ldquo;Conceptual Framework\u0026rdquo; in 3.5; and the same \u0026ldquo;Standardized Training Protocol\u0026rdquo; in 3.6 were utilised.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Results and Discussion","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Comparative Endurance Analysis (Hybrid Vs MAformer)\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the performance of the CNN-LSTM-GRU (Hybrid) and MAformer architectures across 5 forecast horizons. Their performance in terms of Goodness of Fit and horizon decay is also presented in Appendices A and B. A distinct \"performance crossover\" was observed as the lead time extended from the inception of the long-term (168 h) to the extreme of the long-term (240 h) regime. In the short-term regime (24h), the Hybrid model demonstrated superior precision, achieving an R\u003csup\u003e2\u003c/sup\u003e of about 0.892 compared to the MAformer\u0026rsquo;s R\u003csup\u003e2\u003c/sup\u003e of about 0.751. This suggests that for immediate tidal advection, the CNN\u0026rsquo;s local feature extraction combined with the GRU\u0026rsquo;s recurrent memory is more efficient at mapping high-frequency fluctuations. At the 120-hour horizon, the Hybrid model demonstrates its maximum performance advantage, achieving an RMSE that is 22.3% lower than that of the MAformer (0.333 vs. 0.429 respectively). This significant 0.0955 error gap underscores the Hybrid's superior ability to maintain phase alignment during the critical synoptic transition phase (Appendix B)\u0026mdash;a finding consistent with Schmidt et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), who noted that RNN-based hybrids excel when the local hydrodynamic \"memory\" is still physically relevant. However, as the horizon reached the extreme of the long-term regime (240h), the Hybrid model\u0026rsquo;s R\u003csup\u003e2\u003c/sup\u003e decayed significantly to about 0.205 while the RMSE show unprecedented increase to about 0.438. In contrast, the MAformer exhibited higher resilience, outperforming the Hybrid with an R\u003csup\u003e2\u003c/sup\u003e of about 0.241 and a lower RMSE of about 0.405. The MAformer\u0026rsquo;s advantage at 10 days is attributed to its self-attention mechanism, which effectively processed the 241-hour lookback window. While the recurrent units in the Hybrid model likely suffered from information dilution over ~\u0026thinsp;19.4 M2 tidal cycles, the MAformer successfully \"attended\" to periodicities across the entire sequence, capturing the subtidal drift characteristic of the point B in the Outer Shannon Estuary.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMulti-horizon SSS forecast with CNN-LSTM-GRU and MAformer.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOperational Regime\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eForecast Horizon\u003c/p\u003e \u003cp\u003e(h)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePerformance Position\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eShort-term\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e24 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.892024328\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.149914906\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.108157763\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.750893488\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.220369513\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.168943027\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eMedium-term\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e72 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.624070932\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.279050657\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.236711443\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.611834607\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.290250044\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.235846802\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e120 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.518487273\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.333236290\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.276683690\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.485628361\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.428809332\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.361344060\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eLong-term\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e168 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.319393752\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.364305804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.299971077\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.286716687\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.382873008\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.335053076\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e240 Hours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.205662057\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.438079621\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.375534803\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.241104769\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.405137999\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.339359355\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Scale-Dependent Ablation and Predictor Dynamics (Hybrid Vs Maformer)\u003c/h2\u003e \u003cp\u003eThe results of the thematic ablation study (Tables\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Tab9\" class=\"InternalRef\"\u003e9\u003c/span\u003e, and the summary in Table\u0026nbsp;\u003cspan refid=\"Tab10\" class=\"InternalRef\"\u003e10\u003c/span\u003e) reveals that the \u0026ldquo;information value\u0026rdquo; of environmental drivers is not static; it undergoes a fundamental transformation as the operational horizon expands. The comparative ablation heatmap (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) reveals a fundamental transition in predictive logic as the forecast horizon expands from 24 to 240 hours. The colour gradient (dark blue to red) further corroborates this pattern, with deeper blue tones (higher R\u0026sup2; \u0026asymp; 0.75\u0026ndash;0.85) concentrated in the Hybrid panels at shorter horizons, indicating stronger explained variance and forecast skill. Error growth with increasing lead time is evident in both models, reflected by the transition toward yellow\u0026ndash;orange and red tones (R\u0026sup2; \u0026lt; 0.35) at extended horizons. The allied \u0026ldquo;Regime Shift\u0026rdquo; is characterized by a transition from Hydrodynamic Dominance to Climatological Anchor Dependence.\u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e4.2.1 Short-Term Regime (24 Hours): Hydrodynamic Dominance\u003c/h2\u003e \u003cp\u003eIn the short-term regime, both architectures demonstrated high resilience to predictor loss, maintaining R\u003csup\u003e2\u003c/sup\u003e values above 0.85. In this regime, the models function as \u0026ldquo;high-fidelity tidal emulators,\u0026rdquo; where the salinity flux is a direct product of immediate advection. However, the CNN-LSTM-GRU model performed better than the Maformer in response to all the 5 ablation groups. This further validates the relative superiority of the former in such short-term SSS forecasting in the outer estuary. The Hybrid\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e24-hour ablation test between CNN-LSTM-GRU and Maformer.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAblation Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eModel\u0026rsquo;s Performance (R\u003csup\u003e2\u003c/sup\u003e) Position\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eBaseline (Full)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.896933257\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.135205507\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.098653157\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.851888285\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.212425224\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.163956961\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Steric/Mass\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.885161176\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.150311869\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.100331658\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.840144642\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.167355447\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.121138977\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Thermal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.900581700\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.133351443\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.091901315\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.859673946\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.180258449\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.138860242\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Surface\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.878385488\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.175603426\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.145410965\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.795323741\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.222047422\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.169220071\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Local Velocity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.879335760\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.162016699\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.129375088\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.867726632\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.156752105\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.117856269\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003emodel achieved its peak performance (R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.901) when the Thermal group (swpt) was ablated, suggesting that at this scale, sea-water potential temperature acts as \u0026ldquo;climatological noise\u0026rdquo; that distracts the model from high-frequency tidal signals. In terms of predictor sensitivity, the removal of Local Velocity and Surface Forcing (Wind/Pressure/ssh) caused significant R\u003csup\u003e2\u003c/sup\u003e drops, confirming that 24-hour SSS variability in the Outer Shannon Estuary is primarily a function of tidal momentum and surface stress (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e4.2.2 The Medium-Term Regime (72\u0026ndash;120 Hours): The Synoptic Transition\u003c/h2\u003e \u003cp\u003eThe medium-term regime represents the \"stability crossover\" point. During this 3-to-5-day window, the influence of local hydrodynamic \"memory\" begins to fade, and the models must rely on synoptic atmospheric trends. In terms of ablation divergence, at 120h, the Hybrid model began to show increased sensitivity to Surface Forcing, while the MAformer showed a more distributed dependency. In terms of information decay, the first sign of the Hybrid model\u0026rsquo;s (Baseline) performance decay was observed as the R\u003csup\u003e2\u003c/sup\u003e dropped from 0.669 (72h) to 0.585 (120h). This indicates that the recurrent gates are struggling to maintain the phase alignment of tidal cycles over multiple days, marking the beginning of the shift away from deterministic local forcing toward broader environmental trends (Tables\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e and \u003cspan refid=\"Tab7\" class=\"InternalRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e72-hour ablation test between CNN-LSTM-GRU and MAformer.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAblation Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eModel\u0026rsquo;s Performance (R\u003csup\u003e2\u003c/sup\u003e) Position\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eBaseline (Full)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.668731170\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.266150099\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.216379273\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.618443024\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.338101862\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.262775844\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Steric/Mass\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.730026473\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.222844677\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.167225976\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.646609543\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.251102616\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.187637156\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Thermal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.719551413\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.224270387\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.166829226\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.649256718\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.279136094\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.223182635\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Surface\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.715667740\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.264021929\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.215648666\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.598636931\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.272652079\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.204377027\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Local Velocity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.716426165\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.252263786\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.205253971\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.650110004\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.254955983\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.195328534\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e120-hour ablation test between CNN-LSTM-GRU and MAformer.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAblation Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eModel\u0026rsquo;s Performance (R\u003csup\u003e2\u003c/sup\u003e) Position\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eBaseline (Full)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.585352424\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.271818918\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.223405082\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.365146159\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.369032908\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.308609836\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Steric/Mass\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.624782615\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.274110433\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.211652265\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.399055186\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.361192537\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.287648601\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Thermal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.397598025\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.327603148\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.239133078\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.324941169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.387505403\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.309900979\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Surface\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.620563237\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.274679190\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.234284501\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.375974473\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.328057605\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.252043118\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Local Velocity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.583174585\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.266785091\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.209785050\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.369918966\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.344012224\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.269024928\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e4.2.3 Long-Term Regime (168\u0026ndash;240 Hours): Climatological Anchor Dependency\u003c/h2\u003e \u003cp\u003eIn the long-term regime, the predictive logic undergoes a total transformation. The \"memory\" of Local Velocity and short-term Wind stress becomes statistically irrelevant, replaced entirely by oceanic boundary conditions. At the inception of the regime (168 h), the Hybrid model shows relative superiority across all the ablation groups except the Baseline. The regime shift becomes most evident at the 240-hour horizon. Without the Steric/Mass predictors (gmsv, gmmvv), both models show total predictive collapse. Similarly, without the Thermal group predictor (swpt), both models collapsed. However, the No Steric/Mass MAformer\u0026rsquo;s R\u003csup\u003e2\u003c/sup\u003e dropping from 0.172 to a negligible 0.003; and the No Thermal MAformer\u0026rsquo;s R\u003csup\u003e2\u003c/sup\u003e collapsing from 0.102 to 0.007 imply a relatively significant anchor effect (Tables\u0026nbsp;\u003cspan refid=\"Tab8\" class=\"InternalRef\"\u003e8\u003c/span\u003e and \u003cspan refid=\"Tab9\" class=\"InternalRef\"\u003e9\u003c/span\u003e). This proves that at a 10-day horizon, the Outer Shannon\u0026rsquo;s salinity is no longer an estuarine problem but a shelf-sea problem. In terms of architectural de-coherence, this regime highlights the MAformer\u0026rsquo;s strength. Such de-coherence refers to the point where the internal logic of a deep learning model\u0026mdash;specifically the CNN-LSTM-GRU\u0026mdash;mathematically \"unravels\" because it can no longer reconcile the physical relationships between the input predictors and the target salinity at a 240-hour lead time. A de-coherence state could be defined through 3 relevant lenses, the Loss of Temporal Continuity (Recurrent Decay); Stochastic Interference (Noise Overpowering Signal); and the \"Mean-Reversion\" Collapse. While the Hybrid model entered a state of architectural de-coherence\u0026mdash;evidenced by the asterisked (*) relatively high R\u003csup\u003e2\u003c/sup\u003e values with relatively high RMSE and MAE values in the Baseline and No Surface groups (Table\u0026nbsp;\u003cspan refid=\"Tab8\" class=\"InternalRef\"\u003e8\u003c/span\u003e); and the asterisked (**) negative importance values (Table\u0026nbsp;\u003cspan refid=\"Tab11\" class=\"InternalRef\"\u003e11\u003c/span\u003e)\u0026mdash;the MAformer\u0026rsquo;s attention mechanism successfully isolated the \"Climatological Anchors\". By \"attending\" to the global mass signals over the 241-hour window, the MAformer maintained a coherent (though lower) R\u003csup\u003e2\u003c/sup\u003e with relatively low RMSE and MAE values, which imply relatively accurate 240h SSS forecasts (Table\u0026nbsp;\u003cspan refid=\"Tab9\" class=\"InternalRef\"\u003e9\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e168-hour ablation test between CNN-LSTM-GRU and MAformer.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAblation Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eModel\u0026rsquo;s Performance (R\u003csup\u003e2\u003c/sup\u003e) Position\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eBaseline (Full)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.205159791\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.414166785\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.369023338\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.266829291\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.409161521\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.352466184\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Steric/Mass\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.524847594\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.376172051\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.313591343\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.171774763\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.415642084\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.332898032\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Thermal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.208633724\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.417473004\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.321307549\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.102111244\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.491008238\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.408948505\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Surface\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.515116858\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.323842764\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.282772665\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.197248796\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.377983061\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.297896763\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Local Velocity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.470771902\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.340053825\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.280645513\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.172934779\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.404019600\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.331836198\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab8\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 8\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e240-hour ablation test between CNN-LSTM-GRU and MAformer.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAblation Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eModel\u0026rsquo;s Performance (R\u003csup\u003e2\u003c/sup\u003e) Position\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eBaseline (Full)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.219803355*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.518995208\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.438749539\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.215839103\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.447289435\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.352113979\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Steric/Mass\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.023919637\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.473549348\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.388768872\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.002818044\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.433527423\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.377396571\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Thermal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.110696059\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.476712087\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.382465886\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.006665896\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.525376867\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.446542847\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Surface\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.419014486*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.53503734\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.454317608\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.136718881\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.40776054\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.316736389\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eNo Local Velocity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN-LSTM-GRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.35489267\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.402517772\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.333761702\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1st\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.078620458\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.421498015\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.35005752\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2nd\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab9\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 9\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance gap between CNN-LSTM-GRU and MAformer in the 240-Hour Ablation.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHorizon: 240h\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHybrid (Baseline)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMAformer (Baseline)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePerformance Gap\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCorrelation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.220\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.216\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNegligible\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eError\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMAE\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.439\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.352\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eMAformer is 19.8% better\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eError\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eRMSE\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.519\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.447\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eMAformer is 13.8% better\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e4.2.4 Synopsis of the Statistical Behaviors of the Models by Regime\u003c/h2\u003e \u003cp\u003eI. Short-term Regime (24h)\u003c/p\u003e \u003cp\u003eIn the 24-hour window, given that the variables like local velocity, sea surface height, and even wind stress are all physically coupled to this tidal pulse, the two models experience Signal Redundancy. If you remove one group (e.g., Velocity), the model can still \"infer\" the tidal stage from the others. The Hybrid model excels here because its CNN layers are highly efficient at extracting these sharp, repetitive spatial-temporal features from the redundant data stream, leading to its peak performance with R\u003csup\u003e2\u003c/sup\u003e of 0.892\u0026ndash;0.901. The MAformer also benefits, but its Attention heads are designed to look for \"relationships\" rather than just \"patterns.\" In a redundant environment, the MAformer spends computational energy \"attending\" to multiple variables that are essentially telling it the same thing. This is why the Hybrid slightly outperforms the MAformer at the 24h mark\u0026mdash;the Hybrid is a more efficient tool for \"simple\" redundant signals (Table\u0026nbsp;\u003cspan refid=\"Tab10\" class=\"InternalRef\"\u003e10\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eII. Medium-term Regime (72\u0026ndash;120h)\u003c/p\u003e \u003cp\u003eAs we move past 3 days, the \"memory\" of the initial tidal state begins to blur due to cumulative meteorological forcing (wind-driven mixing). This is the Transition / Decay phase. Statistically, we see the R\u003csup\u003e2\u003c/sup\u003e begin to drop (decay) because the relationship between the input and the target is no longer strictly deterministic. The Hybrid model begins to lose its edge as the LSTM/GRU units struggle with the accumulating phase-lag of the tide, while the MAformer begins to show its strength in identifying the longer-period synoptic \"weather\" patterns in the data (Table\u0026nbsp;\u003cspan refid=\"Tab10\" class=\"InternalRef\"\u003e10\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIII. Long-term Regime (168\u0026ndash;240h)\u003c/p\u003e \u003cp\u003eAt the 10-day mark, the \"Redundancy\" has evaporated because the local tides and wind gusts are essentially \"noise\". The wind and velocity no longer correlate with the 10-day salinity trend. The only remaining predictive \"signal\" is the slow-moving, low-frequency ocean state (Steric Mass/Temperature). This is why the models transition from a state of Signal Redundancy (where they have many ways to win) to Anchor Dependency (where they have only one way to win: the Steric/Thermal data). As the models become Climatological Anchor Dependent, if you ablate these specific variables, the R\u003csup\u003e2\u003c/sup\u003e collapses to near zero as shown in the results (Table\u0026nbsp;\u003cspan refid=\"Tab8\" class=\"InternalRef\"\u003e8\u003c/span\u003e). In effect, the MAformer is the better architecture here because its Global Attention mechanism can bypass the \"noise\" of 240 hours of tides to focus exclusively on these \"anchors,\" maintaining a stable (albeit lower) absolute error (MAE) (Table\u0026nbsp;\u003cspan refid=\"Tab10\" class=\"InternalRef\"\u003e10\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab10\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 10\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary of the operational logic for the Hybrid and MAformer.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegime\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrimary Driver\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBetter Architecture\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStatistical Behavior\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eShort-term\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLocal Velocity / Tide\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eHybrid\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSignal Redundancy\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMedium-term\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSurface Wind / Pressure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eHybrid/MAformer\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTransition / Decay\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLong-term\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSteric Mass / Thermal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eMAformer\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eClimatological Anchor Dependent\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Percentage Importance of Predictor Group for the Regime Shift\u003c/h2\u003e \u003cp\u003eTo further delineate the transition from hydrodynamic to climatological regimes, a Percentage Importance analysis was conducted (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e; Table\u0026nbsp;\u003cspan refid=\"Tab11\" class=\"InternalRef\"\u003e11\u003c/span\u003e). At the 24-hour horizon, the importance of all groups remained below 3%, indicating that the Hybrid model is highly robust to individual sensor failures in the short term. Remarkably, the removal of Thermal data at 24h resulted in a negative importance (-0.4%), confirming that thermal signals act as high-frequency noise for daily tidal predictions. However, at the 240-hour horizon, the Steric/Mass group accounted for 98.6% of the MAformer\u0026rsquo;s predictive integrity. This identifies a 'Single Point of Failure' for long-term forecasting. While short-term models are resilient and distribute importance across all 9 predictors, the 10-day forecast is entirely reliant on the global mass-change signal. This quantitative shift justifies the operational recommendation of prioritizing global products, particularly including the GOPAF that was leveraged for this study; and relevant satellite-derived products for strategic estuarine planning.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab11\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 11\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePercentage importance of predictor groups across operational horizons.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOperational Regime\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHorizon\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSteric/Mass (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThermal (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSurface (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eLocal Velocity (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u003cb\u003eShort-term\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e24h\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHybrid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e2.0%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e6.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-1.8%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u003cb\u003eMedium-term\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e120h\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHybrid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e32.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e31.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e-5.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.3%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e25.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e10.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e-3.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-1.3%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u003cb\u003eLong-term\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e240h\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHybrid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e89.1%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e49.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e-90.5%\u003c/b\u003e**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e-61.4%\u003c/b\u003e**\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e98.6%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e96.7%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e36.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e63.4%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eIn Table\u0026nbsp;\u003cspan refid=\"Tab11\" class=\"InternalRef\"\u003e11\u003c/span\u003e, the asterisked (**) negative values (e.g., -90.5% and \u0026minus;\u0026thinsp;61.4%) for the Hybrid architecture at the 240h horizon denote a state of architectural de-coherence. At this extreme lead time, the Hybrid model\u0026rsquo;s recurrent units lose the ability to map long-range dependencies, causing it to overfit to high-frequency residuals in the Surface and Local Velocity groups. The negative importance suggests that the presence of these variables at long horizons introduces 'stochastic interference' for the CNN-LSTM-GRU. This contrasts sharply with the MAformer at the 240h horizon, which maintains positive importance across all groups, demonstrating its superior ability to disentangle noise from signal in the Outer Shannon\u0026rsquo;s complex multi-regime environment.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Operational Synthesis: The \"Information Cross-Over\"\u003c/h2\u003e \u003cp\u003eBy synthesizing both studies, we identify an Information Cross-Over Point. Between 120h and 168h, the models transition from being \"Hydrodynamically Driven\" to \"Climatologically Dependent\" (Table\u0026nbsp;\u003cspan refid=\"Tab11\" class=\"InternalRef\"\u003e11\u003c/span\u003e). This explains why the MAformer, which is designed to identify long-range patterns in complex sequences, begins to take the lead in accuracy as the horizon expands to the extreme of 240h.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Analysis of Metric Divergence: R\u003csup\u003e2\u003c/sup\u003e Stability vs. Error Magnitude\u003c/h2\u003e \u003cp\u003eA notable finding in the 240-hour regime (Table\u0026nbsp;\u003cspan refid=\"Tab8\" class=\"InternalRef\"\u003e8\u003c/span\u003e) is the divergence between the coefficient of determination (R\u003csup\u003e2\u003c/sup\u003e) and the absolute error metrics (RMSE and MAE). In the Baseline (Full) group, both architectures exhibited nearly identical R\u003csup\u003e2\u003c/sup\u003e values (Hybrid: 0.219; MAformer: 0.215), yet their error magnitudes differed substantially. The MAformer achieved a 19.64% lower MAE (0.352 vs. 0.438) and a 13.71% lower RMSE (0.447 vs. 0.518) compared to the Hybrid model. This divergence provides critical insight into how each model handles the increased variance of a 10-day horizon in the Outer Shannon Estuary as subsequently detailed.\u003c/p\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e4.5.1 The \"Mean-Reversion\" of Hybrid Models\u003c/h2\u003e \u003cp\u003eThe Hybrid model\u0026rsquo;s higher error magnitude (RMSE, and MAE) despite its relatively high R\u003csup\u003e2\u003c/sup\u003e suggests that the CNN-LSTM-GRU architecture tends to make larger \"peak-to-peak\" errors. When recurrent units (LSTM/GRU) lose temporal coherence at long horizons, they often revert toward predicting the mean of the training distribution to minimize the global loss function. While this maintains a baseline R\u003csup\u003e2\u003c/sup\u003e (as the model still captures the general trend), it fails to resolve the specific magnitudes of salinity surges or drops, leading to inflated RMSE and MAE.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e4.5.2 The \"Precision Mapping\" of Attention Mechanisms\u003c/h2\u003e \u003cp\u003eIn contrast, the MAformer\u0026rsquo;s lower error metrics indicate a superior ability to resolve the amplitude of salinity fluctuations. Because the Self-Attention mechanism computes dependencies across the entire 241-hour lookback window without the decay associated with recurrent gates, it can identify specific historical \"signatures\" (e.g., a high-spring tide combined with a specific wind stress pattern) that correlate with the future state. This results in a forecast that is more closely aligned with the actual observed values on a point-by-point basis, even when the overall explained variance (R\u003csup\u003e2\u003c/sup\u003e) is constrained by the inherent chaotic nature of the 10-day horizon.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e4.5.3 Physical Implication for the Outer Shannon Estuary\u003c/h2\u003e \u003cp\u003eIn an operational context within the Outer Shannon, this divergence is highly significant. An R\u003csup\u003e2\u003c/sup\u003e of 0.21 tells us that both models are capturing the underlying \"low-frequency\" signal (the Spring-Neap drift). However, the MAformer\u0026rsquo;s significantly lower MAE implies that it is more reliable for determining whether salinity will cross a specific threshold (critical for aquaculture or desalination) at the 10-day mark. The Hybrid model, while identifying the trend, introduces about 20% more uncertainty in the actual magnitude, which could lead to \"false alarms\" or missed environmental hazards.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eThe study has successfully demonstrated the import of a rigorous comparative study of a Hybrid CNN-LSTM-GRU architecture and a MAformer for an extended multi-horizon (24-240h) SSS forecasting in a relatively complex outer part of Shannon, an European estuary. The outputs of the endurance test offer useful operational information that could support decisions of practitioners in their choice between the two models. The detailed results of the systematic ablation across the 5 horizons help to decode the \"information value\" of thermal, steric, surface, and velocity predictors at each horizon. This justifies the need for practitioners to ascertain the scale-dependent importance of environmental predictors in the real-world applications of such DL models. The study also addressed the gap of \"Information Starvation\" by scaling the lookback window to capture up to 19.4 tidal cycles for long-term horizons, a methodology not yet standardized in European estuarine DL research. The result of the endurance test shows the superiority of the Hybrid model for 24\u0026ndash;168 hours forecasts; and the superior resilience of MAformer model for 240 hours forecasts. While both models succumb to the \"Climatological Collapse\" when the anchors are removed, the MAformer demonstrates superior resilience in the Baseline and No Surface groups at 240h. The attention mechanism effectively \"filters\" the 241-hour lookback window to identify long-period subtidal oscillations. Conversely, the Hybrid model suffers from recurrent information decay, where the high-frequency tidal \"noise\" from early in the sequence dilutes the long-term trend signal needed for the 10-day forecast. Overall, the results indicate that the Hybrid model provides superior short- and medium-range stability and greater resilience to input perturbation, while MAformer retains competitive long-range capability under complete predictor configurations but is more sensitive to thematic ablation.\u003c/p\u003e \u003cp\u003eThus, this study has successfully established an operational hierarchy for sensor-efficient estuarine monitoring based on forecast lead times. The relevant thematic findings are synthesized as follows.\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eArchitectural Trade-offs: The CNN-LSTM-GRU Hybrid is the optimal choice for high-precision, short-term (24h) and medium-term (72-120h) operational alerts. However, the MAformer is more robust for strategic, long-term (240h) forecasting.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eRegime Sensitivity: Short-term forecasts are inhibited by the \"noise\" of global thermal/steric variables, whereas long-term forecasts are entirely dependent on them.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWindowing Efficacy: The L\u0026thinsp;=\u0026thinsp;h + 1 scaling strategy proved essential, allowing the MAformer to resolve the nearly 20 M2 cycles required to stabilize a 10-day forecast.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe Shannon Factor: In the outer Shannon Estuary, the transition from local tidal forcing to open-ocean steric influence occurs between the 3-day and 5-day marks.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"6. Recommendations and Future Work","content":"\u003cp\u003e\u003cem\u003e6.1 Operational Recommendations\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eI. \u0026nbsp; \u0026nbsp;Dual-Model Deployment: It is recommended that maritime authorities in the Shannon region adopt a \"Hierarchical Forecasting System\": a Hybrid model for both daily 24h navigation safety and 72-120h early warning information/preparedness; and an MAformer for 10-day environmental management.\u003c/p\u003e\n\u003cp\u003eII. \u0026nbsp;Sensor Prioritization: Investment should be prioritized for high-frequency velocity sensors for daily operations, but real-time integration with global sea-level mass data (Steric/Mass) is mandatory for any long-term planning tools.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e6.2 Future Study Proposals\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eI. \u0026nbsp; \u0026nbsp;Extreme Event Analysis: Future work should investigate how these architectures perform during extreme storm surges or \"1-in-50-year\" fluvial flood events in the Shannon.\u003c/p\u003e\n\u003cp\u003eII. \u0026nbsp;Transfer Learning: A study should be conducted to see if the \"Climatological Anchor\" discovered here for the Shannon can be transferred to other European estuaries (e.g., the Elbe or Tagus) to reduce training time.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e6.3 Study Limitations: Operational Boundaries and Architectural De-coherence\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eWhile this study establishes the MAformer as a resilient tool for long-term forecasting, it is essential to define the boundaries within which these findings remain valid. Every data-driven model possesses a \"prediction horizon limit\" beyond which physical-mathematical alignment fails.\u003c/p\u003e\n\u003cp\u003eI. The De-coherence Threshold\u003c/p\u003e\n\u003cp\u003eAs identified in subsections 4.2.3 and 4.3, the CNN-LSTM-GRU Hybrid reaches a state of architectural de-coherence at the 240-hour mark. This represents a fundamental limitation: recurrent architectures are constrained by their \"memory bottleneck.\" Researchers should be cautioned that extending such models beyond the 10-day limit in macrotidal environments like the Outer Shannon Estuary may result in \"stochastic interference,\" where the model produces outputs based on residual noise rather than actual hydrodynamic drivers.\u003c/p\u003e\n\u003cp\u003eII. Dependency on the \"Climatological Anchor\"\u003c/p\u003e\n\u003cp\u003eThe primary limitation of the MAformer—and indeed any model operating at the 240h time scale—is its absolute dependency on the Steric/Mass (gmsv) and Thermal (swpt) groups. As the ablation study proved, these variables act as the \"Climatological Anchor\". However, the constraint is that if satellite-derived steric data or deep-water temperature sensors experience a telemetry failure, the model's predictive power for horizons \u0026gt;120h will collapse entirely. The model does not have a \"physics-backup\" to compensate for the loss of these specific global signals.\u003c/p\u003e\n\u003cp\u003eIII. Spatial Specificity of the \"Outer\" Estuary\u003c/p\u003e\n\u003cp\u003eThis study specifically targeted the Outer Shannon, where salinity is dominated by open-ocean exchange and shelf-sea interactions; particularly at the point where the AFWA is relatively high. However, there is a constraint. These findings may not be directly transferable to the Inner Shannon (river-dominated) without recalibration. In the inner estuary, river discharge (fluvial forcing) likely replaces Steric/Mass as the primary \"anchor\" for long-term horizons. Therefore, the architectural success of the MAformer here is tied to its ability to process oceanic signals, and its performance under river-dominant regimes remains an area for future validation.\u003c/p\u003e\n\u003cp\u003eIV. Data Density and Frequency\u003c/p\u003e\n\u003cp\u003eThe L=h+1 windowing strategy requires high-density, uninterrupted time-series data. In many operational settings, \"data gaps\" are common. The current architectures are sensitive to these gaps; the MAformer, in particular, requires a complete 241-hour sequence to calculate its Attention Heads effectively. The limitation lies in the \"data-hungry\" nature of Transformer-based models compared to simpler statistical regressions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions:\u003c/strong\u003e Conceptualization, O.A-J.; methodology, O.A-J.; software, O.A-J.; validation, O.A-J.; formal analysis, O.A-J.; investigation, O.A-J.; resources, O.A-J.; data curation, O.A-J. (data downloaded from EU CMEMS repository); writing—original draft preparation, O.A-J.; writing—review and editing, O.A-J. and O.A-J.; visualization, O.A-J.; supervision, O.A-J.; project administration, O.A-J. The author has read and agreed to the published version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e This research received no external funding.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed Consent Statement:\u003c/strong\u003e Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement:\u003c/strong\u003e The dataset used was obtained from the EU CMEMS via\u003c/p\u003e\n\u003cp\u003ehttps://data.marine.copernicus.eu/product/GLOBAL_ANALYSISFORECAST_PHY_001_024/download?dataset=cmems_mod_glo_phy_anfc_0.083deg_PT1H-m_202406 (Data);\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ehttps://data.marine.copernicus.eu/product/GLOBAL_ANALYSISFORECAST_PHY_001_024/download?dataset=cmems_mod_glo_phy_anfc_merged-sl_PT1H-i_202411 (Data);\u003c/p\u003e\n\u003cp\u003ehttps://data.marine.copernicus.eu/product/GLOBAL_ANALYSISFORECAST_PHY_001_024/download?dataset=cmems_mod_glo_phy_anfc_merged-uv_PT1H-i_202211 (Data); and\u003c/p\u003e\n\u003cp\u003ehttps://data.marine.copernicus.eu/product/GLOBAL_ANALYSISFORECAST_PHY_001_024/description or https://doi.org/10.48670/moi-00016 (Metadata documentation).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of Interest:\u003c/strong\u003e The authors declare no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Trial Number:\u003c/strong\u003e Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics and Consent to Participate Declarations:\u003c/strong\u003e Not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAjibola-James, O. (2026). Investigation of High-Frequency Dynamics of Sea Surface Salinity in the Outer Shannon Estuary Using Numerical Model-Derived Data. \u003cem\u003ePreprints\u003c/em\u003e. https://doi.org/10.20944/preprints202601.1029.v1 \u003c/li\u003e\n\u003cli\u003eCho, K., Van Merri\u0026euml;nboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., \u0026amp; Bengio, Y. (2014). Learning phrase representations using RNN encoder-decoder for statistical machine translation. \u003cem\u003earXiv preprint arXiv:1406.1078\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eCostello, M. J. (2025). Ecological responses to salinity variability in Irish estuarine systems. \u003cem\u003eEstuarine, Coastal and Shelf Science, 312\u003c/em\u003e, 108945.\u003c/li\u003e\n\u003cli\u003eGorski, G., Cook, S. E., Snyder, A. M., Appling, A. P., Thompson, T. P., Smith, J. D., Warner, J. C., \u0026amp; Topp, S. N. (2024). Deep learning of estuary salinity dynamics is physically accurate at a fraction of hydrodynamic model computational cost. \u003cem\u003eLimnology and Oceanography\u003c/em\u003e, 69(5), 1070\u0026ndash;1085. https://doi.org/10.1002/lno.12549\u003c/li\u003e\n\u003cli\u003eHochreiter, S., \u0026amp; Schmidhuber, J. (1997). Long short-term memory. \u003cem\u003eNeural Computation\u003c/em\u003e, 9(8), 1735-1780.\u003c/li\u003e\n\u003cli\u003eNguyen, H., et al. (2025). Estuary salinity prediction using machine learning: Case study in the Hau Estuary in Mekong River. \u003cem\u003eWater Supply\u003c/em\u003e. https://doi.org/10.2166/ws.2025.007\u003c/li\u003e\n\u003cli\u003eNing, P., Zhang, C., Zhang, X., \u0026amp; Jiang, X. (2021). Short- to medium-term sea surface height and salinity prediction using an optimized simple recurrent unit deep network. \u003cem\u003eFrontiers in Marine Science\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e, 672280.\u003c/li\u003e\n\u003cli\u003eO\u0026apos;Donncha, F., Grant, J., \u0026amp; Hill, R. J. (2020). An environmental monitoring and forecasting system for sustainable marine aquaculture. \u003cem\u003eJournal of Marine Systems\u003c/em\u003e, 201, 103239. https://doi.org/10.1016/j.jmarsys.2019.103239\u003c/li\u003e\n\u003cli\u003ePritchard, D. W. (1967). What is an estuary: physical viewpoint. \u003cem\u003eEstuaries\u003c/em\u003e, 83, 3-5.\u003c/li\u003e\n\u003cli\u003eQi, S., He, M., Hoang, R., Zhou, Y., Namadi, P., Tom, B., Sandhu, P., Bai, Z., Chung, F., \u0026amp; Ding, Z. (2023). Salinity modeling using deep learning with data augmentation and transfer learning. \u003cem\u003eWater\u003c/em\u003e, 15(13), 2482. https://doi.org/10.3390/w15132482\u003c/li\u003e\n\u003cli\u003eSaccotelli, G., et al. (2024). Enhancing estuary salinity prediction: A machine learning and deep learning based approach. \u003cem\u003eApplied Computing and Geosciences\u003c/em\u003e, 23, 100173. https://doi.org/10.1016/j.acags.2024.100173\u003c/li\u003e\n\u003cli\u003eSchmidt, A., Ziegeler, M., \u0026amp; Mayer, B. (2021). Hybrid CNN-LSTM networks for Sea Surface Salinity in the German Bight. \u003cem\u003eOcean Modelling\u003c/em\u003e, 162, 101804.\u003c/li\u003e\n\u003cli\u003eZhang, Y., et al. (2025). Predictions of saltwater intrusion in the Changjiang Estuary: Integrating machine learning methods with FVCOM. \u003cem\u003eJournal of Hydrology\u003c/em\u003e, 653, 132739. https://doi.org/10.1016/j.jhydrol.2025.132739\u003c/li\u003e\n\u003cli\u003eZheng, R., Sun, Z., Jiao, J., Ma, Q., \u0026amp; Zhao, L. (2024). Salinity prediction based on improved LSTM model in the Qiantang Estuary. \u003cem\u003eJournal of Marine Science and Engineering\u003c/em\u003e, 12(8), 1339. https://doi.org/10.3390/jmse12081339\u003c/li\u003e\n\u003cli\u003eZhu, B., Wang, T., De Meester, J., \u0026amp; Willems, P. (2024). Comparative analysis with statistical and machine learning for modeling salinity along the Scheldt Estuary. \u003cem\u003eWater\u003c/em\u003e, 16(15), 2150. https://doi.org/10.3390/w16152150\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Outer Shannon Estuary, SSS Forecasting, MAformer vs CNN-LSTM-GRU Ablation Study, Architectural De-coherence, Multi-Regime Modelling","lastPublishedDoi":"10.21203/rs.3.rs-9178680/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9178680/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAccurate Sea Surface Salinity (SSS) forecasting is critical for operational management in macrotidal environments, yet the transition between local tidal forcing and long-term climatological drivers remains poorly understood in deep learning. This study presents a multi-regime evaluation and comparative ablation analysis of a CNN-LSTM-GRU (Hybrid) model versus a Multi-Head Attention Transformer (MAformer) in the Outer Shannon Estuary. Leveraging nine robust predictors categorized into Hydrodynamic, Atmospheric, and Steric/Thermal groups, an \"endurance test\" across horizons from 24 to 240 hours was conducted. To ensure physical consistency, a scaled lookback strategy (L\u0026thinsp;=\u0026thinsp;h + 1) was implemented, providing models with up to 19.4 M2 tidal cycles of historical context. Results demonstrate a distinct architectural crossover. The Hybrid model provides superior short- and medium-range stability (24h R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.892 and 120h R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.518), yet reaches architectural de-coherence at 240 hours, characterized by a performance collapse (R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.206). Conversely, the MAformer exhibited superior long-term resilience, achieving lower error magnitudes (RMSE: 0.405 vs. 0.438) at the 10-day horizon. The thematic ablation reveals a scale-dependent regime shift: short-term forecasts are dominated by local velocity signals, whereas 240h stability is entirely dependent on global Steric/Mass and Thermal \"Climatological Anchors\". Without these anchors at 240h, both architectures experience total predictive failure; notably, the MAformer\u0026rsquo;s R\u003csup\u003e2\u003c/sup\u003e dropped from ~\u0026thinsp;0.216 to 0.003. Findings suggest that operational estuarine systems should adopt a hierarchical modelling approach: deploying Hybrid units for daily navigational safety and Attention-based architectures for long-term strategic planning to maintain physical-mathematical alignment across expanding temporal scales.\u003c/p\u003e","manuscriptTitle":"A Comparative Ablation Study of CNN-LSTM-GRU and MAformer Architectures for Operational Multi-regime Salinity Forecasting in the Outer Shannon Estuary","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-23 13:27:25","doi":"10.21203/rs.3.rs-9178680/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3d79f149-7cc2-41fd-9d89-85ff589af93b","owner":[],"postedDate":"March 23rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":64979169,"name":"Artificial Intelligence and Machine Learning"},{"id":64979170,"name":"Physical Geography"},{"id":64979171,"name":"Oceanography"},{"id":64979172,"name":"Hydrology"}],"tags":[],"updatedAt":"2026-03-23T13:27:25+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-23 13:27:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9178680","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9178680","identity":"rs-9178680","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}