Machine Learning–Enhanced Detection of Climate Regime Shifts in the West African Sahel

Machine Learning–Enhanced Detection of Climate Regime Shifts in the West African Sahel | 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 Machine Learning–Enhanced Detection of Climate Regime Shifts in the West African Sahel Mbayang THIAM, Abdoulaye FATY, Awa NIANG This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9182342/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 10 You are reading this latest preprint version Abstract Climate breakpoint detection is fundamental to understanding regime shifts in hydroclimatic systems, yet the dominant statistical tests (Pettitt, Buishand, and SNHT) assume single change-points and offer limited diagnostic capacity. This study tests a Machine Learning (ML) Ensemble framework combining Random Forest and Gradient Boosting classifiers trained on 13 engineered temporal features (CUSUM, rolling divergence statistics, local trend slopes, and rank-based indicators) to detect and characterize multiple breakpoints in climate time series. The approach is applied to monthly temperature and precipitation records from 12 stations spanning Senegal’s diverse climatic zones (1975-2025), alongside classical tests and kernel-based changepoint methods (PELT, Binary Segmentation). Results reveal a pronounced and statistically robust thermal breakpoint concentrated around 1994-1996 across coastal and northern Sahel stations, with a secondary warming shift circa 2010-2015 in inland stations. Cohen’s d effect sizes range from 0.99 to 2.59, confirming large-magnitude warming shifts of +0.5°C to +1.4°C. Precipitation breakpoints are substantially weaker, consistent with high Sahelian rainfall variability. The ML Ensemble method demonstrates superior multi-breakpoint detection capacity and provides continuous probability surfaces rather than binary outcomes, enabling richer uncertainty quantification. These findings carry direct implications for climate adaptation planning, water governance, and territorial resilience strategies across the West African Sahel. climate breakpoint detection machine learning Pettitt test Sahel changepoint analysis temperature trends CUSUM random forest Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1. Introduction The detection of abrupt changes, or breakpoints, in hydroclimatic time series is a central task of climate variability analysis, with significant implications for water resource management, agricultural planning, and territorial resilience. In the West African Sahel, where climate variability is among the highest globally, identifying the timing and magnitude of regime shifts is essential for evidence-based adaptation strategies (Nicholson 2013 ; Descroix et al. 2018 ; Lebel and Ali 2009 ; Biasutti 2019 ). The region experienced a well-documented drought from the late 1960s through the 1980s, followed by partial rainfall recovery, a trajectory whose structural inflection points remain debated in the literature. Classical statistical methods for breakpoint detection, principally the Pettitt test (Pettitt 1979 ), the Buishand range test (Buishand 1982 ), and the Standard Normal Homogeneity Test (SNHT) (Alexandersson 1986 ) have been the standard tools for hydroclimatic homogeneity analysis for over four decades (see Akinsanola and Ogunjobi 2015 , for a recent application in West Africa). These tests, while robust under their assumptions (Reeves et al. 2007 ), share two well-known limitations: (i) they assume a single change-point in the series, and (ii) they produce binary outcomes (significant or not) without quantifying detection uncertainty. The Hubert segmentation procedure (Hubert et al. 1989 ), widely used in francophone West African climatology (Lubès-Niel et al. 1998 ), permits multiple breakpoints but relies on parametric assumptions that may not be satisfied for the heavy-tailed distributions typical of Sahelian rainfall. Recent advances in machine learning (ML) and computational statistics offer promising alternatives. Kernel-based changepoint detection algorithms (PELT (Pruned Exact Linear Time; Killick et al. 2012 ), Binary Segmentation, and window-based methods) can identify multiple breakpoints with computational efficiency. In parallel, supervised and ensemble ML classifiers can be trained on engineered temporal features to produce continuous breakpoint probability surfaces, enabling graduated uncertainty quantification and multi-scale detection. Yet applications of these methods to African climate data remain limited, despite the region’s high vulnerability to climate variability. The present study contributes to filling this gap by developing and applying an ML Ensemble breakpoint detection framework (combining Random Forest and Gradient Boosting classifiers trained on 13 engineered temporal features) to monthly climate records from 12 stations of the Agence Nationale de l’Aviation Civile et de la Météorologie (ANACIM) spanning Senegal’s diverse ecoclimatic zones. Three objectives guide the analysis: (1) to systematically compare classical statistical tests with ML-enhanced methods across both temperature and precipitation variables; (2) to characterize the spatial and temporal structure of detected breakpoints across Senegal; and (3) to evaluate the added diagnostic value of ML approaches for Sahelian climate analysis. 2. Study Area and Data 2.1 Study Area Senegal spans approximately 12.3°N to 16.7°N latitude and 11.3°W to 17.5°W longitude, encompassing a marked north-south climatic gradient from the semi-arid Sahel (Podor, Matam, Linguère: annual rainfall < 400 mm) through the Sudano-Sahelian zone (Akinsanola and Ogunjobi 2017 ) (Kaolack, Diourbel, Tambacounda: 400–800 mm) to the Sudano-Guinean domain (Kolda, Ziguinchor, Kédougou: 800–1500 mm). The coastal stations of Dakar-Yoff, Saint-Louis, and Cap Skirring are additionally modulated by Atlantic maritime influences. This gradient makes Senegal a particularly suitable study area for comparative breakpoint analysis across contrasting hydroclimatic regimes. 2.2 Data Sources and Preprocessing Monthly climate records were obtained from the ANACIM open database (InfoClimat archives) for 12 synoptic stations: Cap Skirring (1988–2025), and Dakar-Yoff, Diourbel, Kaolack, Kédougou, Kolda, Linguère, Matam, Podor, Saint-Louis, Tambacounda, and Ziguinchor (all 1975–2025). Variables include monthly mean minimum temperature (Tmin), monthly mean maximum temperature (Tmax), and monthly cumulative precipitation. Annual aggregates were computed as mean annual temperature Tmean = (Tmin + Tmax)/2, and annual total precipitation. Only years with ≥ 10 months of valid data were retained, yielding 51 complete annual records for 11 stations and 38 for Cap Skirring. Table 1 Summary of the 12 stations used in this study. Station Coordinates Period Years Climatic Zone Cap Skirring 12.39°N, 16.75°W 1988–2025 38 Sudano-Guinean / Coastal Dakar-Yoff 14.74°N, 17.49°W 1975–2025 51 Sahelian / Coastal Diourbel 14.65°N, 16.23°W 1975–2025 51 Sudano-Sahelian Kaolack 14.15°N, 16.08°W 1975–2025 51 Sudano-Sahelian Kédougou 12.56°N, 12.22°W 1975–2025 51 Sudano-Guinean Kolda 12.88°N, 14.97°W 1975–2025 51 Sudano-Guinean Linguère 15.39°N, 15.11°W 1975–2025 51 Sahelian Matam 15.65°N, 13.25°W 1975–2025 51 Sahelian Podor 16.65°N, 14.97°W 1975–2025 51 Sahelian Saint-Louis 16.05°N, 16.50°W 1975–2025 51 Sahelian / Coastal Tambacounda 13.77°N, 13.68°W 1975–2025 51 Sudano-Sahelian Ziguinchor 12.56°N, 16.27°W 1975–2025 51 Sudano-Guinean 3. Methods 3.1 Classical Breakpoint Tests Three classical non-parametric tests were implemented. The Pettitt test (Pettitt 1979 ) detects a single change-point in the mean by maximizing the Mann-Whitney statistic U across all possible partition points. The test statistic K = max|S(t)| follows an approximate distribution under the null hypothesis of homogeneity, with p-value computed as p = 2exp(− 6K²/(n³+n²)). The Buishand range test (Buishand 1982 ) examines the rescaled cumulative departures from the mean, with the test statistic Q* = max|S(t)|/(σ√n). The Standard Normal Homogeneity Test (SNHT; Alexandersson 1986 ) computes T0 = max[t·z̄¹² + (n − t)·z̄²²], where z̄¹ and z̄² are the standardized sub-sample means before and after time t. All tests were evaluated at α = 0.05. 3.2 Kernel-Based Changepoint Methods Three algorithms from the ruptures Python library (Truong et al. 2020 ) were applied. PELT (Pruned Exact Linear Time; Killick et al. 2012 ) minimizes a penalized cost function using dynamic programming with pruning, employing a radial basis function (RBF) kernel and a BIC-derived penalty (λ = ln(n)·σ²). Binary Segmentation (BinSeg) recursively partitions the series by selecting the split that maximizes the L2 cost reduction, with the number of breakpoints set heuristically at n/15. Window-based detection slides a window of width w = 10 years across the series, computing local cost discrepancies, with breakpoints identified where the discrepancy exceeds the BIC penalty. A minimum segment size of 5 years was enforced across all three methods. 3.3 ML Ensemble Breakpoint Detection This study proposes an ML Ensemble framework that transforms breakpoint detection from a binary hypothesis test into a continuous probability estimation problem. The approach proceeds in five stages: Stage 1: Feature Engineering. For each time step t in the annual series, 13 temporal features are computed: (i) raw value; (ii-iii) rolling mean before and after t (window = 5 years); (iv-v) rolling standard deviation before and after t; (vi) absolute mean difference between pre- and post-windows; (vii) standard deviation ratio; (viii-ix) CUSUM and absolute CUSUM; (x) local trend slope via linear regression in a ± 5-year window; (xi-xii) rank and normalized rank; (xiii) proportion of preceding values less than the current value. This feature space captures the statistical signatures of both abrupt shifts and gradual transitions. Stage 2: Pseudo-Label Generation. Classical breakpoints (Pettitt, Buishand, SNHT) and PELT detections serve as pseudo-labels: a ± 2-year neighbourhood around each detected breakpoint is labelled as positive (breakpoint zone), with all remaining time steps labelled negative. This semi-supervised strategy leverages classical expertise while allowing ML refinement. Stage 3: Ensemble Training. Two classifiers are trained on the standardized feature matrix: a Random Forest (Breiman 2001 ; 200 trees, max depth 5, balanced class weights) and a Gradient Boosting classifier (Friedman 2001 ; 100 estimators, max depth 3). Both are implemented via scikit-learn (Pedregosa et al. 2011 ). Stage 4: Probability Surface Construction. The predicted class probabilities from both classifiers are averaged to produce a continuous breakpoint probability surface P(t) ∈ [0, 1] for each time step. This surface quantifies detection confidence without imposing a binary threshold. Stage 5: Peak Detection. Local peaks in P(t) exceeding 0.4, with a minimum inter-peak distance of 5 years, are identified as ML-detected breakpoints. Feature importances from the Random Forest provide interpretability regarding which temporal signatures drive detection. 3.4 Validation and Effect Size Analysis Detected breakpoints were validated through: (i) Welch’s independent samples t-test comparing pre- and post-breakpoint means; (ii) Cohen’s d effect size, computed as d = (µafter − µbefore) / √[(σ²before + σ²after)/2], interpreted as small (0.2), medium (0.5), or large (0.8) following Cohen ( 1988 ); and (iii) inter-method consensus, quantified as the number of methods (out of 5: Pettitt, Buishand, SNHT, PELT, ML Ensemble) detecting a breakpoint within ± 3 years of the consensus year. 4. Results 4.1 Temperature Breakpoints Temperature breakpoint detection yielded highly consistent results across methods and stations (Fig. 2 ). All 12 stations exhibited at least one significant breakpoint detected by either Pettitt or Buishand, with p-values reaching 10⁻¹¹ at Ziguinchor and 10⁻⁸ at Dakar-Yoff and Kaolack. Two distinct temporal clusters emerge: Cluster I (1994–1996) : A primary warming breakpoint concentrated around 1994–1996 is detected at coastal and northern stations (Dakar-Yoff: 1995; Saint-Louis: 1995; Podor: 1995; Linguère: 1995; Ziguinchor: 1996; Kolda: 1994; Cap Skirring: 1994). The ML Ensemble confirms these detections with high probability peaks (P(t) > 0.6) and refines the exact timing to within ± 1 year of classical estimates. This cluster coincides with the widely documented mid-1990s warming acceleration over West Africa associated with the positive phase of the Atlantic Multidecadal Oscillation (AMO; Ting et al. 2009 ). Cluster II (2010–2015) : A secondary warming shift is detected at inland stations further from maritime moderation: Matam (2010), Kaolack (2010), Diourbel (2015), and Tambacounda (2013–2015). The ML Ensemble provides particularly strong diagnostic value here, identifying this secondary shift through its multi-breakpoint capacity. PELT confirms two-phase warming at Kolda (1980, 1995, 2010), Kaolack (1990, 2010), and Tambacounda (1990, 2015), where classical single-breakpoint tests can only report the dominant shift. 4.2 Precipitation Breakpoints In contrast to temperature, precipitation breakpoints were largely non-significant across the network (Fig. 4 ). Only two stations yielded statistically significant detections: Kolda (Pettitt: 1985, p = 0.004; Buishand: 1985, p < 0.001) and Tambacounda (Pettitt: 1996, p = 0.033). The ML Ensemble confirmed the Kolda breakpoint (1984) but identified no additional precipitation breakpoints at other stations. This asymmetry between temperature and precipitation has important ecological and hydrological significance: it suggests that while temperature has undergone discrete regime shifts, rainfall variability in the Sahel remains dominated by interannual and decadal oscillations without clear structural discontinuities, a finding with important implications for drought characterization and water resource planning (see also Ogunrinde et al. 2021 ). 4.3 Method Consensus and Comparative Performance The consensus heatmap (Fig. 5 ) reveals strong inter-method agreement for temperature breakpoints. At Dakar-Yoff, for example, Pettitt (1996), Buishand (1996), PELT (1995), and ML Ensemble (1995) converge within a 1-year window, yielding a 4/5 consensus. Similar convergence is observed at Ziguinchor (4/5), Kolda (4/5), and Saint-Louis (4/5). The SNHT consistently fails to reach significance at α = 0.05 despite identifying breakpoint years within 1–2 years of the Pettitt/Buishand consensus, suggesting a power limitation relative to the other tests for this sample size. The ML Ensemble provides unique added value in two respects. First, it detects secondary breakpoints that single-breakpoint classical tests cannot identify (e.g., the 2010–2015 warming shift at Kaolack, Kolda, and Tambacounda). Second, its continuous probability output (Fig. 3 ) enables uncertainty-aware interpretation: high, narrow peaks indicate sharp regime shifts, while broad, moderate peaks suggest gradual transitions, a distinction invisible to binary hypothesis tests. 4.4 Effect Size and Statistical Validation The dumbbell plot and Cohen’s d analysis (Fig. 6 ) provide quantitative validation of temperature breakpoints. All 12 stations exhibit positive warming shifts, with pre-to-post breakpoint temperature increases ranging from + 0.34°C (Kédougou) to + 1.36°C (Diourbel). Effect sizes are uniformly large: Cohen’s d ranges from 0.99 (Saint-Louis) to 2.59 (Diourbel), with 11 of 12 stations exceeding the large-effect threshold of 0.8. Welch’s t-tests confirm significance at p < 0.05 for 11 of 12 stations, with Kédougou as the sole exception (p = 0.47), likely attributable to its shorter post-breakpoint period. 4.5 Feature Importance and CUSUM Diagnostics The Random Forest feature importance analysis (Fig. 7 a) reveals that CUSUM-derived features dominate the ML Ensemble’s discriminative capacity, with absolute CUSUM ranking highest, followed by normalized rank and the mean difference between pre- and post-windows. This finding is consistent with theoretical expectations: the CUSUM statistic is the cumulative signature of persistent departures from the series mean, making it a natural indicator of regime shifts. The local trend slope, by contrast, ranks lower, suggesting that the ML approach prioritizes cumulative divergence signals over instantaneous gradient information. The CUSUM diagnostic plot (Fig. 7 b) visually confirms the detected breakpoints: CUSUM curves for all six displayed stations transition from negative (cooler-than-average phase) to positive (warmer-than-average phase) trajectories, with the inflection points corresponding closely to the ML-detected breakpoints (marked by dots). The slope change is especially pronounced at Ziguinchor and Dakar-Yoff, consistent with their large effect sizes. 4.6 Spatial Synthesis The spatial distribution of breakpoints (Fig. 8 ) reveals a coherent geographic structure. For temperature, the dominant 1994–1996 breakpoint is ubiquitous but most strongly expressed at coastal stations modulated by Atlantic SST forcing, consistent with AMO-mediated warming. The secondary 2010–2015 breakpoint is concentrated at inland stations, suggesting amplified continental warming independent of maritime buffering. Consensus strength (bubble size) is highest at Ziguinchor, Dakar-Yoff, and Kolda (4–5 methods agreeing), confirming these as the most robustly documented warming transitions. For precipitation, the map is dominated by non-significant detections, with only Kolda and Tambacounda showing meaningful breakpoints, both in the wetter southern zone. 5. Discussion 5.1 Methodological Advances The ML Ensemble framework introduced here advances breakpoint detection methodology in three key respects. First, it replaces binary hypothesis testing with continuous probability estimation, providing a richer representation of detection uncertainty. This is particularly valuable for climate series where breakpoints may be gradual rather than abrupt, a common situation in temperature trends but poorly captured by classical tests that assume step-function changes. Second, the semi-supervised pseudo-labelling strategy effectively bridges classical and ML approaches: rather than discarding decades of established methodology, it leverages classical detections as informative priors while allowing ML classifiers to refine, confirm, or reject these initial estimates. Third, the feature importance analysis provides scientific interpretability, a critical requirement for climate research that purely algorithmic approaches (e.g., neural networks) often lack (Molnar 2020 ). The dominance of CUSUM features in the Random Forest importance ranking is noteworthy. It suggests that the most discriminative signal for breakpoint detection is the cumulative, persistent departure from historical norms, the type of signal that the CUSUM statistic was designed to capture. The ML framework, without being explicitly informed of the CUSUM’s theoretical significance, independently identifies this feature as the most informative. This convergence lends mutual validation, as classical CUSUM theory is corroborated by data-driven feature selection. 5.2 Climatic Interpretation The mid-1990s warming breakpoint identified across most Senegalese stations is consistent with the broader literature on West African temperature trends. Fontaine et al. ( 2011 ) documented accelerated warming over the Sahel beginning in the 1990s, linked to positive AMO phases and enhanced greenhouse forcing (see also Sylla et al. 2016 ). The spatial pattern observed here, stronger expression at coastal stations, supports the hypothesis that Atlantic SST variability plays a modulating role, likely through its influence on the West African Monsoon circulation and associated heat transport. The secondary 2010–2015 breakpoint at inland stations is a notable finding of this study, enabled by the multi-breakpoint capacity of the ML Ensemble and PELT methods. This shift coincides with a period of record-breaking global temperatures and may reflect the superposition of continued greenhouse warming on decadal variability. The inland amplification is consistent with known physical mechanisms: continental interiors, lacking the thermal buffering of oceanic influence, are expected to respond more strongly to radiative forcing changes. The near-absence of significant precipitation breakpoints is also instructive. While the Sahel drought of the 1970s-1980s is well documented, the recovery since the 1990s has been spatially heterogeneous and not a simple reversal (Panthou et al. 2018 ; Lebel and Ali 2009 ). The high interannual variability of Sahelian rainfall, with coefficients of variation exceeding 30% at northern stations, likely masks any structural breakpoint that classical tests or ML methods could detect with 50 years of data. This finding cautions against interpreting the absence of a detected breakpoint as the absence of change; rather, it reflects the limitations of all breakpoint methods when signal-to-noise ratios are low. 5.3 Implications for Water Governance and Territorial Planning The confirmed + 0.5°C to + 1.4°C warming shifts carry direct implications for water governance in the Senegal River Basin and coastal zones (Bodian et al. 2018 ). Elevated temperatures increase evapotranspiration, reduce effective rainfall, and amplify water stress on rain-fed agriculture, the dominant livelihood in the Sudano-Sahelian zone. The spatial differentiation of breakpoint timing has practical significance for adaptation planning: coastal municipalities experienced the warming shift two decades earlier than inland communes, suggesting that adaptation measures implemented at the coast may inform future strategies for the interior. These findings support the integration of breakpoint-informed climate diagnostics into territorial planning frameworks, particularly for the Plan National d’Adaptation (PNA) and commune-level Plans Climat Territoriaux. 5.4 Limitations Several limitations merit acknowledgment. First, the pseudo-labelling strategy introduces circular dependency: the ML Ensemble is partially trained on classical detections, potentially inheriting their biases. Future work could employ fully unsupervised deep learning approaches (e.g., autoencoders for anomaly detection) to achieve independence. Second, the feature engineering window of 5 years is a tuneable hyperparameter whose optimization was not systematically explored. Third, the ANACIM records, while among the longest in West Africa, remain relatively short (51 years) for robust detection of multi-decadal breakpoints. Fourth, the p-value approximations for the classical tests are asymptotic and may be imprecise for the sample sizes used here; permutation-based alternatives could improve inferential rigour. Finally, the analysis focused on annual aggregates; monthly or seasonal breakpoint detection could reveal additional structure, particularly in the onset and cessation dates of the rainy season (Mendes and Fragoso 2025 ). 6. Conclusion This study demonstrates that Machine Learning-enhanced breakpoint detection provides meaningful diagnostic advantages over classical statistical tests for Sahelian climate analysis. The ML Ensemble framework combining Random Forest and Gradient Boosting classifiers trained on 13 engineered temporal features confirms and refines classical breakpoint estimates while enabling multi-breakpoint detection and continuous probability quantification. Applied to 12 ANACIM stations across Senegal (1975–2025), the analysis reveals: A spatially coherent, statistically robust thermal breakpoint around 1994–1996 at coastal and northern stations, with Cohen’s d values of 0.99–2.59 (uniformly large effects) and warming magnitudes of + 0.5°C to + 1.4°C. A secondary inland warming shift circa 2010–2015 is detectable only through multi-breakpoint methods (ML Ensemble and PELT), underscoring the limitations of classical single-breakpoint tests. Precipitation breakpoints remain largely non-significant, reflecting the dominance of interannual variability over structural discontinuities in Sahelian rainfall. The ML Ensemble’s feature importance analysis converges independently on CUSUM-derived metrics as the most discriminative breakpoint indicators, providing mutual validation between classical theory and data-driven inference. The continuous probability surface produced by the ensemble offers a more nuanced representation of breakpoint certainty than the binary outcomes of traditional tests. These findings carry practical implications for climate adaptation governance in Senegal and the broader Sahel, where projected warming trajectories are among the steepest globally (Dosio et al. 2021 ; Oguntunde et al. 2024 ). The integration of ML-enhanced climate diagnostics into territorial planning frameworks, particularly for water-energy-food-ecosystem nexus governance, could strengthen evidence-based decision-making at local, national, and transboundary scales. Future work should explore deep learning architectures, seasonal-resolution analysis, and extension to the full ECOWAS station network to consolidate and generalize these findings. Statements and Declarations Funding The authors did not receive support from any organization for the submitted work. Competing Interests The authors have no relevant financial or non-financial interests to disclose. Author Contributions Mbayang Thiam was responsible for the study conception and design, data collection and analysis, methodology development, and manuscript preparation. Abdoulaye Faty helped with data collection, analysis methodology and manuscript preparation. Pr Awa Niang supervised the overall study and contributed to the manuscript preparation. Data Availability The climate datasets analysed in this study were obtained from the Agence Nationale de l’Aviation Civile et de la Météorologie (ANACIM) open database (InfoClimat archives). The processed datasets generated during this study are available from the corresponding author on reasonable request. Use of AI Tools AI-assisted tools were used during manuscript preparation for language editing, reference formatting, and compliance with journal style guidelines. All scientific content, data analysis, interpretation, and conclusions are entirely the work of the author, who assumes full responsibility for the integrity of the manuscript. References Akinsanola AA, Ogunjobi KO (2015) Recent homogeneity analysis and long-term spatio-temporal rainfall trends in Nigeria. Theor Appl Climatol 128(1-2):275-289. Akinsanola AA, Ogunjobi KO (2017) Evaluation of rainfall simulations over West Africa in dynamically downscaled CMIP5 global circulation models. Theor Appl Climatol 128(1-2):201-218. Alexandersson H (1986) A homogeneity test applied to precipitation data. J Climatol 6(6):661-675. Aminikhanghahi S, Cook DJ (2017) A survey of methods for time series change point detection. Knowl Inf Syst 51(2):339-367. Biasutti M (2019) Rainfall trends in the African Sahel: Characteristics, processes, and causes. WIREs Clim Change, 10(4), e591. Bodian A, Dezetter A, Deme A, Diop L (2018) Hydrological evaluation of TRMM rainfall over the upper Senegal River basin. Hydrology 5(1):1-18. Breiman L (2001) Random forests. Mach Learn 45(1):5-32. Buishand TA (1982) Some methods for testing the homogeneity of rainfall records. J Hydrol 58(1-2):11-27. Chalapathy R, Chawla S (2019) Deep learning for anomaly detection: A survey. arXiv preprint arXiv:1901.03407. Cohen J (1988) Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates. Descroix L, Diongue Niang A, Panthou G, Bodian A, Sane Y, Dacosta H, ..., Vischel T (2018) Evolution récente de la pluviométrie en Afrique de l’Ouest. Climatologie 15:1-30. Dosio A, Jury MW, Almazroui M, Ashfaq M, Diallo I, Giorgi F, ..., Tamoffo AT (2021) Projected future daily characteristics of African precipitation based on global (CMIP5, CMIP6) and regional (CORDEX, CORDEX-CORE) climate models. Clim Dyn 57(11-12):3135-3158. Fontaine B, Janicot S, Monerie PA (2011) Recent changes in air temperature, heat waves occurrences, and atmospheric circulation in Northern Africa. J Geophys Res, 116(D1). Friedman JH (2001) Greedy function approximation: A gradient boosting machine. Ann Stat 29(5):1189-1232. Hubert P, Carbonnel JP, Chaouche A (1989) Segmentation des séries hydrométéorologiques: Application à des séries de précipitations et de débits de l’Afrique de l’Ouest. J Hydrol 110(3-4):349-367. Killick R, Fearnhead P, Eckley IA (2012) Optimal detection of changepoints with a linear computational cost. J Am Stat Assoc 107(500):1590-1598. Lebel T, Ali A (2009) Recent trends in the Central and Western Sahel rainfall regime (1990-2007). J Hydrol 375(1-2):52-64. Lubès-Niel H, Masson JM, Paturel JE, Servat E (1998) Variabilité climatique et statistiques: étude par simulation de la puissance et de la robustesse de quelques tests utilisés pour vérifier l’homogénéité de chroniques. Revue des Sciences de l’Eau 11(3):383-408. Mendes DS, Fragoso M (2025) Variability and trends of the rainy season in West Africa with a special focus on Guinea-Bissau. Theor Appl Climatol, 156, 103. Molnar C (2020) Interpretable Mach Learn: A Guide for Making Black Box Models Explainable. Lulu.com. Nicholson SE (2013) The West African Sahel: A review of recent studies on the rainfall regime and its interannual variability. ISRN Meteorol 2013, 1-32. Ogunrinde AT, Oguntunde PG, Akinwumiju AS, Fasinmirin JT (2021) Characterization of drought using four drought indices under climate change in the Sahel region of Nigeria. Theor Appl Climatol 143:843-860. Oguntunde PG, Abiodun BJ, Lischeid G (2024) Precipitation, temperature and potential evapotranspiration for 1991-2020 climate normals over Africa. Theor Appl Climatol 155:5385-5406. Panthou G, Lebel T, Vischel T, Quantin G, Sane Y, Ba A, ..., Diongue Niang A (2018) Rainfall intensification in tropical semi-arid regions: the Sahelian case. Environ Res Lett 13(6):064013. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, ..., Duchesnay É (2011) Scikit-learn: Machine learning in Python. J Mach Learn Res 12:2825-2830. Pettitt AN (1979) A non-parametric approach to the change-point problem. J R Stat Soc C Appl Stat 28(2):126-135. Reeves J, Chen J, Wang XL, Lund R, Lu QQ (2007) A review and comparison of changepoint detection techniques for climate data. J Appl Meteorol Climatol 46(6):900-915. Sylla MB, Nikiema PM, Gibba P, Kebe I, Klutse NAB (2016) Climate change over West Africa: Recent trends and future projections. In Adaptation to Climate Change and Variability in Rural West Africa (pp. 25-40). Springer. Ting M, Kushnir Y, Seager R, Li C (2009) Forced and internal twentieth-century SST trends in the North Atlantic. J Clim 22(6):1469-1481. Truong C, Oudre L, Vayer N (2020) Selective review of offline change point detection methods. Signal Process, 167, 107299. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 13 May, 2026 Reviews received at journal 27 Apr, 2026 Reviews received at journal 26 Apr, 2026 Reviewers agreed at journal 14 Apr, 2026 Reviewers agreed at journal 12 Apr, 2026 Reviewers agreed at journal 31 Mar, 2026 Reviewers invited by journal 26 Mar, 2026 Editor assigned by journal 22 Mar, 2026 Submission checks completed at journal 22 Mar, 2026 First submitted to journal 20 Mar, 2026 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-9182342","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":613156601,"identity":"76af9255-573f-4d43-9767-83fbd5c82dd9","order_by":0,"name":"Mbayang THIAM","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABEElEQVRIiWNgGAWjYDACCTiDsYGBoQLESmBg4CFGCw9YyxnStAAJxjYitOjObn4m8XNHLYO9dHPjp5vzDsvztycwPnjbxmBvjkOL2Z1jZpK9Z44z8MgcbJbO3XbYcMaZB8yGc9sYEnc24NByI8FMgrftGNBhiQ0gLQkMNxLYpHnbGBIMDuDSkv5N8i9ES/Pv3DmHE+RvJLD/Bmqxx60lxwxoZg1IS5t0bsPhBAOgLcxALYwbcGm5c6bYWrbtAA/PjcQ265xj6YYbzzxslpxzTiIRp5bb7Rtvvm2rk2Ofkf74dk6Ntbzc8eSDH96U2eB0GBCwAGPkMHJEgOIUkSqwAeYPDAx1+BSMglEwCkbBSAcA+7BfEYnJWOYAAAAASUVORK5CYII=","orcid":"","institution":"Institut de Gouvernance Territoriale et de Développement Local (IGTDL), Université Cheikh Anta Diop (UCAD)","correspondingAuthor":true,"prefix":"","firstName":"Mbayang","middleName":"","lastName":"THIAM","suffix":""},{"id":613156602,"identity":"21048895-33d4-40bc-a404-d22f2800faf5","order_by":1,"name":"Abdoulaye FATY","email":"","orcid":"","institution":"Cheikh Anta Diop University","correspondingAuthor":false,"prefix":"","firstName":"Abdoulaye","middleName":"","lastName":"FATY","suffix":""},{"id":613156603,"identity":"99a91b1f-a3ca-4c80-b3a5-9341ff5c460c","order_by":2,"name":"Awa NIANG","email":"","orcid":"","institution":"Cheikh Anta Diop University","correspondingAuthor":false,"prefix":"","firstName":"Awa","middleName":"","lastName":"NIANG","suffix":""}],"badges":[],"createdAt":"2026-03-21 00:08:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9182342/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9182342/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105817284,"identity":"83749e47-ac8e-4b16-9648-361b33c28eb3","added_by":"auto","created_at":"2026-03-31 12:30:31","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":244221,"visible":true,"origin":"","legend":"\u003cp\u003eStudy area and the ANACIM synoptic station network across Senegal.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/5ed49ed5cec55961c83131d2.png"},{"id":105904855,"identity":"7d8672c7-a50b-4bfa-a171-3cf02bca98c1","added_by":"auto","created_at":"2026-04-01 10:10:50","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":289437,"visible":true,"origin":"","legend":"\u003cp\u003eMean annual temperature time series with detected breakpoints across 12 Senegalese stations (1975-2025).\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/be8b2f434ec01ea318237998.png"},{"id":105817285,"identity":"313ae5af-3c62-4799-add5-6c6b1a7179d2","added_by":"auto","created_at":"2026-03-31 12:30:31","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":265848,"visible":true,"origin":"","legend":"\u003cp\u003eML Ensemble breakpoint probability surfaces for six representative stations.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/a8e724f79866ef3a83a9cbd6.png"},{"id":105817290,"identity":"8d538f97-3c75-49e1-af4a-92aa3eabfa0b","added_by":"auto","created_at":"2026-03-31 12:30:31","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":219329,"visible":true,"origin":"","legend":"\u003cp\u003eAnnual total precipitation with detected breakpoints. Bar heights represent annual totals. Significant breakpoints (Pettitt, Buishand) are marked at Kolda and Tambacounda only.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/14c327853e7498d518168ca3.png"},{"id":105817287,"identity":"5adb2468-c703-4914-aa66-c78606f687d5","added_by":"auto","created_at":"2026-03-31 12:30:31","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":127444,"visible":true,"origin":"","legend":"\u003cp\u003eBreakpoint detection consensus matrix for mean annual temperature.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/8b445c73cc30c406f0176ed3.png"},{"id":105817288,"identity":"9dd7d6e1-ca69-4682-a37d-8a0f5b1a89b5","added_by":"auto","created_at":"2026-03-31 12:30:31","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":86124,"visible":true,"origin":"","legend":"\u003cp\u003eValidation of temperature breakpoints. (a) Dumbbell plot of mean temperature before (blue) and after (red) the primary breakpoint. (b) Cohen’s d effect sizes. (c) Statistical significance expressed as −log₁₀(p-value).\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/3d1e06ee1a96ee8e50c9ad5e.png"},{"id":105904121,"identity":"297c64a0-e81a-4ffd-964a-98c38e2c3f83","added_by":"auto","created_at":"2026-04-01 10:04:34","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":112334,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Mean Random Forest feature importance across all stations. (b) CUSUM diagnostic curves for six representative stations.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/df6720f51eb7606667f8f59a.png"},{"id":105904326,"identity":"fdcef907-7f9b-48f0-bce5-f0cdf1a81cbf","added_by":"auto","created_at":"2026-04-01 10:07:29","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":101374,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial distribution of climate breakpoints across Senegal.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/d0ac22c215cbb44d20ee2610.png"},{"id":106401589,"identity":"2c5cc571-c3c6-4e0a-998e-cc1a0dffe1a5","added_by":"auto","created_at":"2026-04-08 09:07:42","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1976531,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9182342/v1/cec0dca2-6a5f-4393-8792-bfa7d5f7ef4b.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Machine Learning–Enhanced Detection of Climate Regime Shifts in the West African Sahel","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe detection of abrupt changes, or breakpoints, in hydroclimatic time series is a central task of climate variability analysis, with significant implications for water resource management, agricultural planning, and territorial resilience. In the West African Sahel, where climate variability is among the highest globally, identifying the timing and magnitude of regime shifts is essential for evidence-based adaptation strategies (Nicholson \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Descroix et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Lebel and Ali \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Biasutti \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The region experienced a well-documented drought from the late 1960s through the 1980s, followed by partial rainfall recovery, a trajectory whose structural inflection points remain debated in the literature.\u003c/p\u003e \u003cp\u003eClassical statistical methods for breakpoint detection, principally the Pettitt test (Pettitt \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e1979\u003c/span\u003e), the Buishand range test (Buishand \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1982\u003c/span\u003e), and the Standard Normal Homogeneity Test (SNHT) (Alexandersson \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e1986\u003c/span\u003e) have been the standard tools for hydroclimatic homogeneity analysis for over four decades (see Akinsanola and Ogunjobi \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2015\u003c/span\u003e, for a recent application in West Africa). These tests, while robust under their assumptions (Reeves et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2007\u003c/span\u003e), share two well-known limitations: (i) they assume a single change-point in the series, and (ii) they produce binary outcomes (significant or not) without quantifying detection uncertainty. The Hubert segmentation procedure (Hubert et al. \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e1989\u003c/span\u003e), widely used in francophone West African climatology (Lub\u0026egrave;s-Niel et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e1998\u003c/span\u003e), permits multiple breakpoints but relies on parametric assumptions that may not be satisfied for the heavy-tailed distributions typical of Sahelian rainfall.\u003c/p\u003e \u003cp\u003eRecent advances in machine learning (ML) and computational statistics offer promising alternatives. Kernel-based changepoint detection algorithms (PELT (Pruned Exact Linear Time; Killick et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2012\u003c/span\u003e), Binary Segmentation, and window-based methods) can identify multiple breakpoints with computational efficiency. In parallel, supervised and ensemble ML classifiers can be trained on engineered temporal features to produce continuous breakpoint probability surfaces, enabling graduated uncertainty quantification and multi-scale detection. Yet applications of these methods to African climate data remain limited, despite the region\u0026rsquo;s high vulnerability to climate variability.\u003c/p\u003e \u003cp\u003eThe present study contributes to filling this gap by developing and applying an ML Ensemble breakpoint detection framework (combining Random Forest and Gradient Boosting classifiers trained on 13 engineered temporal features) to monthly climate records from 12 stations of the Agence Nationale de l\u0026rsquo;Aviation Civile et de la M\u0026eacute;t\u0026eacute;orologie (ANACIM) spanning Senegal\u0026rsquo;s diverse ecoclimatic zones. Three objectives guide the analysis: (1) to systematically compare classical statistical tests with ML-enhanced methods across both temperature and precipitation variables; (2) to characterize the spatial and temporal structure of detected breakpoints across Senegal; and (3) to evaluate the added diagnostic value of ML approaches for Sahelian climate analysis.\u003c/p\u003e"},{"header":"2. Study Area and Data","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study Area\u003c/h2\u003e \u003cp\u003eSenegal spans approximately 12.3\u0026deg;N to 16.7\u0026deg;N latitude and 11.3\u0026deg;W to 17.5\u0026deg;W longitude, encompassing a marked north-south climatic gradient from the semi-arid Sahel (Podor, Matam, Lingu\u0026egrave;re: annual rainfall\u0026thinsp;\u0026lt;\u0026thinsp;400 mm) through the Sudano-Sahelian zone (Akinsanola and Ogunjobi \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) (Kaolack, Diourbel, Tambacounda: 400\u0026ndash;800 mm) to the Sudano-Guinean domain (Kolda, Ziguinchor, K\u0026eacute;dougou: 800\u0026ndash;1500 mm). The coastal stations of Dakar-Yoff, Saint-Louis, and Cap Skirring are additionally modulated by Atlantic maritime influences. This gradient makes Senegal a particularly suitable study area for comparative breakpoint analysis across contrasting hydroclimatic regimes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Data Sources and Preprocessing\u003c/h2\u003e \u003cp\u003eMonthly climate records were obtained from the ANACIM open database (InfoClimat archives) for 12 synoptic stations: Cap Skirring (1988\u0026ndash;2025), and Dakar-Yoff, Diourbel, Kaolack, K\u0026eacute;dougou, Kolda, Lingu\u0026egrave;re, Matam, Podor, Saint-Louis, Tambacounda, and Ziguinchor (all 1975\u0026ndash;2025). Variables include monthly mean minimum temperature (Tmin), monthly mean maximum temperature (Tmax), and monthly cumulative precipitation. Annual aggregates were computed as mean annual temperature Tmean = (Tmin\u0026thinsp;+\u0026thinsp;Tmax)/2, and annual total precipitation. Only years with \u0026ge;\u0026thinsp;10 months of valid data were retained, yielding 51 complete annual records for 11 stations and 38 for Cap Skirring.\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\u003eSummary of the 12 stations used in this study.\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=\"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\u003eStation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCoordinates\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePeriod\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eYears\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eClimatic Zone\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCap Skirring\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.39\u0026deg;N, 16.75\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1988\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSudano-Guinean / Coastal\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDakar-Yoff\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.74\u0026deg;N, 17.49\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSahelian / Coastal\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiourbel\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.65\u0026deg;N, 16.23\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSudano-Sahelian\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKaolack\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.15\u0026deg;N, 16.08\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSudano-Sahelian\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eK\u0026eacute;dougou\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.56\u0026deg;N, 12.22\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSudano-Guinean\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKolda\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.88\u0026deg;N, 14.97\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSudano-Guinean\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLingu\u0026egrave;re\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15.39\u0026deg;N, 15.11\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSahelian\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMatam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15.65\u0026deg;N, 13.25\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSahelian\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePodor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16.65\u0026deg;N, 14.97\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSahelian\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSaint-Louis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16.05\u0026deg;N, 16.50\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSahelian / Coastal\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTambacounda\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13.77\u0026deg;N, 13.68\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSudano-Sahelian\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZiguinchor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.56\u0026deg;N, 16.27\u0026deg;W\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1975\u0026ndash;2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSudano-Guinean\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"},{"header":"3. Methods","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Classical Breakpoint Tests\u003c/h2\u003e \u003cp\u003eThree classical non-parametric tests were implemented. The Pettitt test (Pettitt \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e1979\u003c/span\u003e) detects a single change-point in the mean by maximizing the Mann-Whitney statistic U across all possible partition points. The test statistic K\u0026thinsp;=\u0026thinsp;max|S(t)| follows an approximate distribution under the null hypothesis of homogeneity, with p-value computed as p\u0026thinsp;=\u0026thinsp;2exp(\u0026minus;\u0026thinsp;6K\u0026sup2;/(n\u0026sup3;+n\u0026sup2;)). The Buishand range test (Buishand \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1982\u003c/span\u003e) examines the rescaled cumulative departures from the mean, with the test statistic Q* = max|S(t)|/(σ\u0026radic;n). The Standard Normal Homogeneity Test (SNHT; Alexandersson \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e1986\u003c/span\u003e) computes T0\u0026thinsp;=\u0026thinsp;max[t\u0026middot;z̄\u0026sup1;\u0026sup2; + (n\u0026thinsp;\u0026minus;\u0026thinsp;t)\u0026middot;z̄\u0026sup2;\u0026sup2;], where z̄\u0026sup1; and z̄\u0026sup2; are the standardized sub-sample means before and after time t. All tests were evaluated at α\u0026thinsp;=\u0026thinsp;0.05.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Kernel-Based Changepoint Methods\u003c/h2\u003e \u003cp\u003eThree algorithms from the \u003cem\u003eruptures\u003c/em\u003e Python library (Truong et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) were applied. PELT (Pruned Exact Linear Time; Killick et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) minimizes a penalized cost function using dynamic programming with pruning, employing a radial basis function (RBF) kernel and a BIC-derived penalty (λ\u0026thinsp;=\u0026thinsp;ln(n)\u0026middot;σ\u0026sup2;). Binary Segmentation (BinSeg) recursively partitions the series by selecting the split that maximizes the L2 cost reduction, with the number of breakpoints set heuristically at n/15. Window-based detection slides a window of width w\u0026thinsp;=\u0026thinsp;10 years across the series, computing local cost discrepancies, with breakpoints identified where the discrepancy exceeds the BIC penalty. A minimum segment size of 5 years was enforced across all three methods.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.3 ML Ensemble Breakpoint Detection\u003c/h2\u003e \u003cp\u003eThis study proposes an ML Ensemble framework that transforms breakpoint detection from a binary hypothesis test into a continuous probability estimation problem. The approach proceeds in five stages:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eStage 1: Feature Engineering. For each time step t in the annual series, 13 temporal features are computed: (i) raw value; (ii-iii) rolling mean before and after t (window\u0026thinsp;=\u0026thinsp;5 years); (iv-v) rolling standard deviation before and after t; (vi) absolute mean difference between pre- and post-windows; (vii) standard deviation ratio; (viii-ix) CUSUM and absolute CUSUM; (x) local trend slope via linear regression in a\u0026thinsp;\u0026plusmn;\u0026thinsp;5-year window; (xi-xii) rank and normalized rank; (xiii) proportion of preceding values less than the current value. This feature space captures the statistical signatures of both abrupt shifts and gradual transitions.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eStage 2: Pseudo-Label Generation. Classical breakpoints (Pettitt, Buishand, SNHT) and PELT detections serve as pseudo-labels: a\u0026thinsp;\u0026plusmn;\u0026thinsp;2-year neighbourhood around each detected breakpoint is labelled as positive (breakpoint zone), with all remaining time steps labelled negative. This semi-supervised strategy leverages classical expertise while allowing ML refinement.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eStage 3: Ensemble Training. Two classifiers are trained on the standardized feature matrix: a Random Forest (Breiman \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2001\u003c/span\u003e; 200 trees, max depth 5, balanced class weights) and a Gradient Boosting classifier (Friedman \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2001\u003c/span\u003e; 100 estimators, max depth 3). Both are implemented via scikit-learn (Pedregosa et al. \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2011\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eStage 4: Probability Surface Construction. The predicted class probabilities from both classifiers are averaged to produce a continuous breakpoint probability surface P(t) \u0026isin; [0, 1] for each time step. This surface quantifies detection confidence without imposing a binary threshold.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eStage 5: Peak Detection. Local peaks in P(t) exceeding 0.4, with a minimum inter-peak distance of 5 years, are identified as ML-detected breakpoints. Feature importances from the Random Forest provide interpretability regarding which temporal signatures drive detection.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Validation and Effect Size Analysis\u003c/h2\u003e \u003cp\u003eDetected breakpoints were validated through: (i) Welch\u0026rsquo;s independent samples t-test comparing pre- and post-breakpoint means; (ii) Cohen\u0026rsquo;s \u003cem\u003ed\u003c/em\u003e effect size, computed as d = (\u0026micro;after\u0026thinsp;\u0026minus;\u0026thinsp;\u0026micro;before) / \u0026radic;[(σ\u0026sup2;before\u0026thinsp;+\u0026thinsp;σ\u0026sup2;after)/2], interpreted as small (0.2), medium (0.5), or large (0.8) following Cohen (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e1988\u003c/span\u003e); and (iii) inter-method consensus, quantified as the number of methods (out of 5: Pettitt, Buishand, SNHT, PELT, ML Ensemble) detecting a breakpoint within \u0026plusmn;\u0026thinsp;3 years of the consensus year.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Temperature Breakpoints\u003c/h2\u003e \u003cp\u003eTemperature breakpoint detection yielded highly consistent results across methods and stations (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). All 12 stations exhibited at least one significant breakpoint detected by either Pettitt or Buishand, with p-values reaching 10⁻\u0026sup1;\u0026sup1; at Ziguinchor and 10⁻⁸ at Dakar-Yoff and Kaolack. Two distinct temporal clusters emerge:\u003c/p\u003e \u003cp\u003e \u003cb\u003eCluster I (1994\u0026ndash;1996)\u003c/b\u003e: A primary warming breakpoint concentrated around 1994\u0026ndash;1996 is detected at coastal and northern stations (Dakar-Yoff: 1995; Saint-Louis: 1995; Podor: 1995; Lingu\u0026egrave;re: 1995; Ziguinchor: 1996; Kolda: 1994; Cap Skirring: 1994). The ML Ensemble confirms these detections with high probability peaks (P(t)\u0026thinsp;\u0026gt;\u0026thinsp;0.6) and refines the exact timing to within \u0026plusmn;\u0026thinsp;1 year of classical estimates. This cluster coincides with the widely documented mid-1990s warming acceleration over West Africa associated with the positive phase of the Atlantic Multidecadal Oscillation (AMO; Ting et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2009\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cb\u003eCluster II (2010\u0026ndash;2015)\u003c/b\u003e: A secondary warming shift is detected at inland stations further from maritime moderation: Matam (2010), Kaolack (2010), Diourbel (2015), and Tambacounda (2013\u0026ndash;2015). The ML Ensemble provides particularly strong diagnostic value here, identifying this secondary shift through its multi-breakpoint capacity. PELT confirms two-phase warming at Kolda (1980, 1995, 2010), Kaolack (1990, 2010), and Tambacounda (1990, 2015), where classical single-breakpoint tests can only report the dominant shift.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Precipitation Breakpoints\u003c/h2\u003e \u003cp\u003eIn contrast to temperature, precipitation breakpoints were largely non-significant across the network (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Only two stations yielded statistically significant detections: Kolda (Pettitt: 1985, p\u0026thinsp;=\u0026thinsp;0.004; Buishand: 1985, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and Tambacounda (Pettitt: 1996, p\u0026thinsp;=\u0026thinsp;0.033). The ML Ensemble confirmed the Kolda breakpoint (1984) but identified no additional precipitation breakpoints at other stations. This asymmetry between temperature and precipitation has important ecological and hydrological significance: it suggests that while temperature has undergone discrete regime shifts, rainfall variability in the Sahel remains dominated by interannual and decadal oscillations without clear structural discontinuities, a finding with important implications for drought characterization and water resource planning (see also Ogunrinde et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Method Consensus and Comparative Performance\u003c/h2\u003e \u003cp\u003eThe consensus heatmap (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e) reveals strong inter-method agreement for temperature breakpoints. At Dakar-Yoff, for example, Pettitt (1996), Buishand (1996), PELT (1995), and ML Ensemble (1995) converge within a 1-year window, yielding a 4/5 consensus. Similar convergence is observed at Ziguinchor (4/5), Kolda (4/5), and Saint-Louis (4/5). The SNHT consistently fails to reach significance at α\u0026thinsp;=\u0026thinsp;0.05 despite identifying breakpoint years within 1\u0026ndash;2 years of the Pettitt/Buishand consensus, suggesting a power limitation relative to the other tests for this sample size.\u003c/p\u003e \u003cp\u003eThe ML Ensemble provides unique added value in two respects. First, it detects secondary breakpoints that single-breakpoint classical tests cannot identify (e.g., the 2010\u0026ndash;2015 warming shift at Kaolack, Kolda, and Tambacounda). Second, its continuous probability output (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) enables uncertainty-aware interpretation: high, narrow peaks indicate sharp regime shifts, while broad, moderate peaks suggest gradual transitions, a distinction invisible to binary hypothesis tests.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Effect Size and Statistical Validation\u003c/h2\u003e \u003cp\u003eThe dumbbell plot and Cohen\u0026rsquo;s d analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e) provide quantitative validation of temperature breakpoints. All 12 stations exhibit positive warming shifts, with pre-to-post breakpoint temperature increases ranging from +\u0026thinsp;0.34\u0026deg;C (K\u0026eacute;dougou) to +\u0026thinsp;1.36\u0026deg;C (Diourbel). Effect sizes are uniformly large: Cohen\u0026rsquo;s d ranges from 0.99 (Saint-Louis) to 2.59 (Diourbel), with 11 of 12 stations exceeding the large-effect threshold of 0.8. Welch\u0026rsquo;s t-tests confirm significance at p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 for 11 of 12 stations, with K\u0026eacute;dougou as the sole exception (p\u0026thinsp;=\u0026thinsp;0.47), likely attributable to its shorter post-breakpoint period.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Feature Importance and CUSUM Diagnostics\u003c/h2\u003e \u003cp\u003eThe Random Forest feature importance analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003ea) reveals that CUSUM-derived features dominate the ML Ensemble\u0026rsquo;s discriminative capacity, with absolute CUSUM ranking highest, followed by normalized rank and the mean difference between pre- and post-windows. This finding is consistent with theoretical expectations: the CUSUM statistic is the cumulative signature of persistent departures from the series mean, making it a natural indicator of regime shifts. The local trend slope, by contrast, ranks lower, suggesting that the ML approach prioritizes cumulative divergence signals over instantaneous gradient information.\u003c/p\u003e \u003cp\u003eThe CUSUM diagnostic plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eb) visually confirms the detected breakpoints: CUSUM curves for all six displayed stations transition from negative (cooler-than-average phase) to positive (warmer-than-average phase) trajectories, with the inflection points corresponding closely to the ML-detected breakpoints (marked by dots). The slope change is especially pronounced at Ziguinchor and Dakar-Yoff, consistent with their large effect sizes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.6 Spatial Synthesis\u003c/h2\u003e \u003cp\u003eThe spatial distribution of breakpoints (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e) reveals a coherent geographic structure. For temperature, the dominant 1994\u0026ndash;1996 breakpoint is ubiquitous but most strongly expressed at coastal stations modulated by Atlantic SST forcing, consistent with AMO-mediated warming. The secondary 2010\u0026ndash;2015 breakpoint is concentrated at inland stations, suggesting amplified continental warming independent of maritime buffering. Consensus strength (bubble size) is highest at Ziguinchor, Dakar-Yoff, and Kolda (4\u0026ndash;5 methods agreeing), confirming these as the most robustly documented warming transitions. For precipitation, the map is dominated by non-significant detections, with only Kolda and Tambacounda showing meaningful breakpoints, both in the wetter southern zone.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Discussion","content":"\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Methodological Advances\u003c/h2\u003e \u003cp\u003eThe ML Ensemble framework introduced here advances breakpoint detection methodology in three key respects. First, it replaces binary hypothesis testing with continuous probability estimation, providing a richer representation of detection uncertainty. This is particularly valuable for climate series where breakpoints may be gradual rather than abrupt, a common situation in temperature trends but poorly captured by classical tests that assume step-function changes. Second, the semi-supervised pseudo-labelling strategy effectively bridges classical and ML approaches: rather than discarding decades of established methodology, it leverages classical detections as informative priors while allowing ML classifiers to refine, confirm, or reject these initial estimates. Third, the feature importance analysis provides scientific interpretability, a critical requirement for climate research that purely algorithmic approaches (e.g., neural networks) often lack (Molnar \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe dominance of CUSUM features in the Random Forest importance ranking is noteworthy. It suggests that the most discriminative signal for breakpoint detection is the cumulative, persistent departure from historical norms, the type of signal that the CUSUM statistic was designed to capture. The ML framework, without being explicitly informed of the CUSUM\u0026rsquo;s theoretical significance, independently identifies this feature as the most informative. This convergence lends mutual validation, as classical CUSUM theory is corroborated by data-driven feature selection.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Climatic Interpretation\u003c/h2\u003e \u003cp\u003eThe mid-1990s warming breakpoint identified across most Senegalese stations is consistent with the broader literature on West African temperature trends. Fontaine et al. (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2011\u003c/span\u003e) documented accelerated warming over the Sahel beginning in the 1990s, linked to positive AMO phases and enhanced greenhouse forcing (see also Sylla et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). The spatial pattern observed here, stronger expression at coastal stations, supports the hypothesis that Atlantic SST variability plays a modulating role, likely through its influence on the West African Monsoon circulation and associated heat transport.\u003c/p\u003e \u003cp\u003eThe secondary 2010\u0026ndash;2015 breakpoint at inland stations is a notable finding of this study, enabled by the multi-breakpoint capacity of the ML Ensemble and PELT methods. This shift coincides with a period of record-breaking global temperatures and may reflect the superposition of continued greenhouse warming on decadal variability. The inland amplification is consistent with known physical mechanisms: continental interiors, lacking the thermal buffering of oceanic influence, are expected to respond more strongly to radiative forcing changes.\u003c/p\u003e \u003cp\u003eThe near-absence of significant precipitation breakpoints is also instructive. While the Sahel drought of the 1970s-1980s is well documented, the recovery since the 1990s has been spatially heterogeneous and not a simple reversal (Panthou et al. \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Lebel and Ali \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). The high interannual variability of Sahelian rainfall, with coefficients of variation exceeding 30% at northern stations, likely masks any structural breakpoint that classical tests or ML methods could detect with 50 years of data. This finding cautions against interpreting the absence of a detected breakpoint as the absence of change; rather, it reflects the limitations of all breakpoint methods when signal-to-noise ratios are low.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Implications for Water Governance and Territorial Planning\u003c/h2\u003e \u003cp\u003eThe confirmed\u0026thinsp;+\u0026thinsp;0.5\u0026deg;C to +\u0026thinsp;1.4\u0026deg;C warming shifts carry direct implications for water governance in the Senegal River Basin and coastal zones (Bodian et al. \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Elevated temperatures increase evapotranspiration, reduce effective rainfall, and amplify water stress on rain-fed agriculture, the dominant livelihood in the Sudano-Sahelian zone. The spatial differentiation of breakpoint timing has practical significance for adaptation planning: coastal municipalities experienced the warming shift two decades earlier than inland communes, suggesting that adaptation measures implemented at the coast may inform future strategies for the interior. These findings support the integration of breakpoint-informed climate diagnostics into territorial planning frameworks, particularly for the Plan National d\u0026rsquo;Adaptation (PNA) and commune-level Plans Climat Territoriaux.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e5.4 Limitations\u003c/h2\u003e \u003cp\u003eSeveral limitations merit acknowledgment. First, the pseudo-labelling strategy introduces circular dependency: the ML Ensemble is partially trained on classical detections, potentially inheriting their biases. Future work could employ fully unsupervised deep learning approaches (e.g., autoencoders for anomaly detection) to achieve independence. Second, the feature engineering window of 5 years is a tuneable hyperparameter whose optimization was not systematically explored. Third, the ANACIM records, while among the longest in West Africa, remain relatively short (51 years) for robust detection of multi-decadal breakpoints. Fourth, the p-value approximations for the classical tests are asymptotic and may be imprecise for the sample sizes used here; permutation-based alternatives could improve inferential rigour. Finally, the analysis focused on annual aggregates; monthly or seasonal breakpoint detection could reveal additional structure, particularly in the onset and cessation dates of the rainy season (Mendes and Fragoso \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study demonstrates that Machine Learning-enhanced breakpoint detection provides meaningful diagnostic advantages over classical statistical tests for Sahelian climate analysis. The ML Ensemble framework combining Random Forest and Gradient Boosting classifiers trained on 13 engineered temporal features confirms and refines classical breakpoint estimates while enabling multi-breakpoint detection and continuous probability quantification. Applied to 12 ANACIM stations across Senegal (1975\u0026ndash;2025), the analysis reveals:\u003c/p\u003e \u003cp\u003eA spatially coherent, statistically robust thermal breakpoint around 1994\u0026ndash;1996 at coastal and northern stations, with Cohen\u0026rsquo;s d values of 0.99\u0026ndash;2.59 (uniformly large effects) and warming magnitudes of +\u0026thinsp;0.5\u0026deg;C to +\u0026thinsp;1.4\u0026deg;C. A secondary inland warming shift circa 2010\u0026ndash;2015 is detectable only through multi-breakpoint methods (ML Ensemble and PELT), underscoring the limitations of classical single-breakpoint tests. Precipitation breakpoints remain largely non-significant, reflecting the dominance of interannual variability over structural discontinuities in Sahelian rainfall.\u003c/p\u003e \u003cp\u003eThe ML Ensemble\u0026rsquo;s feature importance analysis converges independently on CUSUM-derived metrics as the most discriminative breakpoint indicators, providing mutual validation between classical theory and data-driven inference. The continuous probability surface produced by the ensemble offers a more nuanced representation of breakpoint certainty than the binary outcomes of traditional tests.\u003c/p\u003e \u003cp\u003eThese findings carry practical implications for climate adaptation governance in Senegal and the broader Sahel, where projected warming trajectories are among the steepest globally (Dosio et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Oguntunde et al. \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The integration of ML-enhanced climate diagnostics into territorial planning frameworks, particularly for water-energy-food-ecosystem nexus governance, could strengthen evidence-based decision-making at local, national, and transboundary scales. Future work should explore deep learning architectures, seasonal-resolution analysis, and extension to the full ECOWAS station network to consolidate and generalize these findings.\u003c/p\u003e"},{"header":"Statements and Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors did not receive support from any organization for the submitted work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMbayang Thiam was responsible for the study conception and design, data collection and analysis, methodology development, and manuscript preparation. Abdoulaye Faty helped with data collection, analysis methodology and manuscript preparation. Pr Awa Niang supervised the overall study and contributed to the manuscript preparation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe climate datasets analysed in this study were obtained from the Agence Nationale de l’Aviation Civile et de la Météorologie (ANACIM) open database (InfoClimat archives). The processed datasets generated during this study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eUse of AI Tools\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAI-assisted tools were used during manuscript preparation for language editing, reference formatting, and compliance with journal style guidelines. All scientific content, data analysis, interpretation, and conclusions are entirely the work of the author, who assumes full responsibility for the integrity of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAkinsanola AA, Ogunjobi KO (2015) Recent homogeneity analysis and long-term spatio-temporal rainfall trends in Nigeria. Theor Appl Climatol 128(1-2):275-289.\u003c/li\u003e\n \u003cli\u003eAkinsanola AA, Ogunjobi KO (2017) Evaluation of rainfall simulations over West Africa in dynamically downscaled CMIP5 global circulation models. Theor Appl Climatol 128(1-2):201-218.\u003c/li\u003e\n \u003cli\u003eAlexandersson H (1986) A homogeneity test applied to precipitation data. J Climatol 6(6):661-675.\u003c/li\u003e\n \u003cli\u003eAminikhanghahi S, Cook DJ (2017) A survey of methods for time series change point detection. Knowl Inf Syst 51(2):339-367.\u003c/li\u003e\n \u003cli\u003eBiasutti M (2019) Rainfall trends in the African Sahel: Characteristics, processes, and causes. WIREs Clim Change, 10(4), e591.\u003c/li\u003e\n \u003cli\u003eBodian A, Dezetter A, Deme A, Diop L (2018) Hydrological evaluation of TRMM rainfall over the upper Senegal River basin. Hydrology 5(1):1-18.\u003c/li\u003e\n \u003cli\u003eBreiman L (2001) Random forests. Mach Learn 45(1):5-32.\u003c/li\u003e\n \u003cli\u003eBuishand TA (1982) Some methods for testing the homogeneity of rainfall records. J Hydrol 58(1-2):11-27.\u003c/li\u003e\n \u003cli\u003eChalapathy R, Chawla S (2019) Deep learning for anomaly detection: A survey. arXiv preprint arXiv:1901.03407.\u003c/li\u003e\n \u003cli\u003eCohen J (1988) Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates.\u003c/li\u003e\n \u003cli\u003eDescroix L, Diongue Niang A, Panthou G, Bodian A, Sane Y, Dacosta H, ..., Vischel T (2018) Evolution r\u0026eacute;cente de la pluviom\u0026eacute;trie en Afrique de l\u0026rsquo;Ouest. Climatologie 15:1-30.\u003c/li\u003e\n \u003cli\u003eDosio A, Jury MW, Almazroui M, Ashfaq M, Diallo I, Giorgi F, ..., Tamoffo AT (2021) Projected future daily characteristics of African precipitation based on global (CMIP5, CMIP6) and regional (CORDEX, CORDEX-CORE) climate models. Clim Dyn 57(11-12):3135-3158.\u003c/li\u003e\n \u003cli\u003eFontaine B, Janicot S, Monerie PA (2011) Recent changes in air temperature, heat waves occurrences, and atmospheric circulation in Northern Africa. J Geophys Res, 116(D1).\u003c/li\u003e\n \u003cli\u003eFriedman JH (2001) Greedy function approximation: A gradient boosting machine. Ann Stat 29(5):1189-1232.\u003c/li\u003e\n \u003cli\u003eHubert P, Carbonnel JP, Chaouche A (1989) Segmentation des s\u0026eacute;ries hydrom\u0026eacute;t\u0026eacute;orologiques: Application \u0026agrave; des s\u0026eacute;ries de pr\u0026eacute;cipitations et de d\u0026eacute;bits de l\u0026rsquo;Afrique de l\u0026rsquo;Ouest. J Hydrol 110(3-4):349-367.\u003c/li\u003e\n \u003cli\u003eKillick R, Fearnhead P, Eckley IA (2012) Optimal detection of changepoints with a linear computational cost. J Am Stat Assoc 107(500):1590-1598.\u003c/li\u003e\n \u003cli\u003eLebel T, Ali A (2009) Recent trends in the Central and Western Sahel rainfall regime (1990-2007). J Hydrol 375(1-2):52-64.\u003c/li\u003e\n \u003cli\u003eLub\u0026egrave;s-Niel H, Masson JM, Paturel JE, Servat E (1998) Variabilit\u0026eacute; climatique et statistiques: \u0026eacute;tude par simulation de la puissance et de la robustesse de quelques tests utilis\u0026eacute;s pour v\u0026eacute;rifier l\u0026rsquo;homog\u0026eacute;n\u0026eacute;it\u0026eacute; de chroniques. Revue des Sciences de l\u0026rsquo;Eau 11(3):383-408.\u003c/li\u003e\n \u003cli\u003eMendes DS, Fragoso M (2025) Variability and trends of the rainy season in West Africa with a special focus on Guinea-Bissau. Theor Appl Climatol, 156, 103.\u003c/li\u003e\n \u003cli\u003eMolnar C (2020) Interpretable Mach Learn: A Guide for Making Black Box Models Explainable. Lulu.com.\u003c/li\u003e\n \u003cli\u003eNicholson SE (2013) The West African Sahel: A review of recent studies on the rainfall regime and its interannual variability. ISRN Meteorol 2013, 1-32.\u003c/li\u003e\n \u003cli\u003eOgunrinde AT, Oguntunde PG, Akinwumiju AS, Fasinmirin JT (2021) Characterization of drought using four drought indices under climate change in the Sahel region of Nigeria. Theor Appl Climatol 143:843-860.\u003c/li\u003e\n \u003cli\u003eOguntunde PG, Abiodun BJ, Lischeid G (2024) Precipitation, temperature and potential evapotranspiration for 1991-2020 climate normals over Africa. Theor Appl Climatol 155:5385-5406.\u003c/li\u003e\n \u003cli\u003ePanthou G, Lebel T, Vischel T, Quantin G, Sane Y, Ba A, ..., Diongue Niang A (2018) Rainfall intensification in tropical semi-arid regions: the Sahelian case. Environ Res Lett 13(6):064013.\u003c/li\u003e\n \u003cli\u003ePedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, ..., Duchesnay \u0026Eacute; (2011) Scikit-learn: Machine learning in Python. J Mach Learn Res 12:2825-2830.\u003c/li\u003e\n \u003cli\u003ePettitt AN (1979) A non-parametric approach to the change-point problem. J R Stat Soc C Appl Stat 28(2):126-135.\u003c/li\u003e\n \u003cli\u003eReeves J, Chen J, Wang XL, Lund R, Lu QQ (2007) A review and comparison of changepoint detection techniques for climate data. J Appl Meteorol Climatol 46(6):900-915.\u003c/li\u003e\n \u003cli\u003eSylla MB, Nikiema PM, Gibba P, Kebe I, Klutse NAB (2016) Climate change over West Africa: Recent trends and future projections. In Adaptation to Climate Change and Variability in Rural West Africa (pp. 25-40). Springer.\u003c/li\u003e\n \u003cli\u003eTing M, Kushnir Y, Seager R, Li C (2009) Forced and internal twentieth-century SST trends in the North Atlantic. J Clim 22(6):1469-1481.\u003c/li\u003e\n \u003cli\u003eTruong C, Oudre L, Vayer N (2020) Selective review of offline change point detection methods. Signal Process, 167, 107299.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"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":"theoretical-and-applied-climatology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"taac","sideBox":"Learn more about [Theoretical and Applied Climatology](https://www.springer.com/journal/704)","snPcode":"704","submissionUrl":"https://submission.nature.com/new-submission/704/3","title":"Theoretical and Applied Climatology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"climate breakpoint detection, machine learning, Pettitt test, Sahel, changepoint analysis, temperature trends, CUSUM, random forest","lastPublishedDoi":"10.21203/rs.3.rs-9182342/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9182342/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eClimate breakpoint detection is fundamental to understanding regime shifts in hydroclimatic systems, yet the dominant statistical tests (Pettitt, Buishand, and SNHT) assume single change-points and offer limited diagnostic capacity. This study tests a Machine Learning (ML) Ensemble framework combining Random Forest and Gradient Boosting classifiers trained on 13 engineered temporal features (CUSUM, rolling divergence statistics, local trend slopes, and rank-based indicators) to detect and characterize multiple breakpoints in climate time series. The approach is applied to monthly temperature and precipitation records from 12 stations spanning Senegal’s diverse climatic zones (1975-2025), alongside classical tests and kernel-based changepoint methods (PELT, Binary Segmentation). Results reveal a pronounced and statistically robust thermal breakpoint concentrated around 1994-1996 across coastal and northern Sahel stations, with a secondary warming shift circa 2010-2015 in inland stations. Cohen’s \u003cem\u003ed\u003c/em\u003e effect sizes range from 0.99 to 2.59, confirming large-magnitude warming shifts of +0.5°C to +1.4°C. Precipitation breakpoints are substantially weaker, consistent with high Sahelian rainfall variability. The ML Ensemble method demonstrates superior multi-breakpoint detection capacity and provides continuous probability surfaces rather than binary outcomes, enabling richer uncertainty quantification. These findings carry direct implications for climate adaptation planning, water governance, and territorial resilience strategies across the West African Sahel.\u003c/p\u003e","manuscriptTitle":"Machine Learning–Enhanced Detection of Climate Regime Shifts in the West African Sahel","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-31 12:30:21","doi":"10.21203/rs.3.rs-9182342/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-05-14T01:58:55+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-27T12:33:46+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-26T09:49:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"26055140878095728988441986091587707771","date":"2026-04-14T13:08:54+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"3762424916730691271085143727832414794","date":"2026-04-12T15:30:54+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"1691358609507006334770004255721478246","date":"2026-03-31T20:40:42+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-27T00:12:18+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-23T00:54:26+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-23T00:54:09+00:00","index":"","fulltext":""},{"type":"submitted","content":"Theoretical and Applied Climatology","date":"2026-03-20T23:57:43+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"theoretical-and-applied-climatology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"taac","sideBox":"Learn more about [Theoretical and Applied Climatology](https://www.springer.com/journal/704)","snPcode":"704","submissionUrl":"https://submission.nature.com/new-submission/704/3","title":"Theoretical and Applied Climatology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"e5914fee-7c3d-454a-8383-9edc3e0b35bc","owner":[],"postedDate":"March 31st, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-14T01:58:55+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-05-14T02:09:21+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-31 12:30:21","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9182342","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9182342","identity":"rs-9182342","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}