Synergistic Effects of Climate and LULC Change on Urban Rainfall–Runoff Dynamics: A Data-Driven Framework With Lstm–Attention Models

Synergistic Effects of Climate and LULC Change on Urban Rainfall–Runoff Dynamics: A Data-Driven Framework With Lstm–Attention Models | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 31 October 2025 V1 Latest version Share on Synergistic Effects of Climate and LULC Change on Urban Rainfall–Runoff Dynamics: A Data-Driven Framework With Lstm–Attention Models Authors : Marcio Schmidt 0000-0003-2716-2360 [email protected] , Carlos Eugênio Pereira , and Alan Petrônio Pinheiro Authors Info & Affiliations https://doi.org/10.22541/au.176192481.11762995/v1 209 views 162 downloads Contents Abstract Supplementary Material Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Flood warning systems are essential tools for mitigating the negative impacts that arise from hydrological disasters, but to be effective these require accurate and detailed information generated from hydrological simulations in order to assess the impact of flash flooding or dam breaks. However, these simulations are performed on current data, while neglecting the dynamics of land use and land cover (LULC) changes in urban areas over time and the effects of climate change on precipitation volume forecasts and increased surface runoff. Here, we propose a data-driven framework that integrates multivariate LSTM networks with Attention Models, Monte Carlo uncertainty analysis, and CA-Markov LULC projections to estimate future rainfall-runoff conditions. Applied to the São Pedro basin in Uberlândia, Brazil, resulted in a projected increase in the flooded area and higher peaks at the outlet compared to 2025, demonstrating the potential increase in risk to the population and urban vulnerability. This combined approach demonstrates a relevant methodological contribution to accurately prediction urban changes with applicability in support decision-making and urban flood risk management. Synergistic Effects of Climate and LULC Change on Urban Rainfall–Runoff Dynamics: A Data-Driven Framework With Lstm–Attention Models Marcio Augusto Reolon Schmidt 1 , Carlos Eugênio Pereira 2 , and Alan Petrônio Pinheiro 3 1 Programa de Pós-Graduação em Engenharia Civil, Faculdade de Engenharia Civil, Universidade Federal de Uberlândia 2 Programa de Pós-Graduação em Engenharia Civil, Faculdade de Engenharia Civil, Universidade Federal de Uberlândia 3 Laboratório de Redes Inteligentes, Faculdade de Engenharia Elétrica, Universidade Federal de Uberlândia Corresponding Author : Bloco 1Y - Av. João Naves de Ávila, 2121 - Santa Mônica, Uberlândia - MG, Brazil, zip-code: 38400-902; [email protected] Abstract (up to 300 words) Flood warning systems are essential tools for mitigating the negative impacts that arise from hydrological disasters, but to be effective these require accurate and detailed information generated from hydrological simulations in order to assess the impact of flash flooding or dam breaks. However, these simulations are performed on current data, while neglecting the dynamics of land use and land cover (LULC) changes in urban areas over time and the effects of climate change on precipitation volume forecasts and increased surface runoff. Here, we propose a data-driven framework that integrates multivariate LSTM networks with Attention Models, Monte Carlo uncertainty analysis, and CA-Markov LULC projections to estimate future rainfall-runoff conditions. Applied to the São Pedro basin in Uberlândia, Brazil, resulted in a projected increase in the flooded area and higher peaks at the outlet compared to 2025, demonstrating the potential increase in risk to the population and urban vulnerability. This combined approach demonstrates a relevant methodological contribution to accurately prediction urban changes with applicability in support decision-making and urban flood risk management. Keywords: Rainfall-Runoff estimation; Climate change; LSTM- Attention Models modelling; Disaster Prevention; Flash Flooding simulation. 1. INTRODUCTION Cities are dynamic environments that are constantly expanding through the cognizant change of natural space into urbanised and impermeable areas. In the context of climate change, there has been a systemic increase in natural disasters, of which floods are the most destructive, posing a significant and growing threat to lives and property worldwide [1, 2] while representing one of the most serious environmental catastrophes, with irreversible ecological, social, infrastructural and economic consequences [3, 4] and a high risk of casualties and property loss [5]. The assessment of flood risk through changes in runoff, associated with hypothetical scenarios of flooding or dam failure, has received considerable attention in recent decades, whether in flood risk management, emergency response, or planning to mitigate its effects [6]. Proactive measures, such as modelling and assessing flood or dam-breaking risks, are essential for the development of evacuation plans, crisis management and assessment of associated damage [7, 8]. Therefore, flood warning systems, although essential, require accurate and detailed information to be effective [9,10]. As dynamic environments, urban basins possess a high percentage of impervious areas, high population concentration and key infrastructure [11, 12]. In light of such, the ability to predict water levels in urban areas is of particular importance due to their impact on urban infrastructure and dams [4,7, 8, 13]. When considering flood hydrographs recorded during extreme weather events [1, 7] an upward trend is noted, when compared to previous years. In urban areas, the risk of the effects of these events puts the population and infrastructure at greater risk [10]. It is for this reason, a need to understand how these phenomena will evolve in the future becomes essential in an attempt to establish patterns of extreme precipitation on a global and regional scale [11, 14]. Huang et al. [15] and Cea & Costabile [16] agree that changes in rainfall patterns contribute toward modifying surface runoff as well as increasing the magnitude of floods. Accurate, high-resolution estimates of surface runoff are essential for effective flood mitigation and water planning [17], although considered a complex task due to its non-linear factors with low adherence to stochastic analyses [15]. Another relevant aspect to consider is the dynamics of land use and occupation, which notably alters parameters such as the Manning coefficient and the consequent speed of flood propagation throughout the river basin. In this context, simulations performed only with current LULC and precipitation data do not adequately predict which areas are most vulnerable to the effects of changes in urban basins, which is essential for developing flood susceptibility assessments, mitigating damage to people and property, and aiding urban planning. This study addresses this gap by proposing the synergistic integration of machine learning models for changes in runoff and flood areas through future scenarios, in a non-linear or purely stochastic manner. Beyond simplistic models such as MLP [18] and LSTM [19, 20, 21], this research proposes the incorporation of data characteristics and attention models in the parameterisation of historical precipitation series, and the use of LULC neighbourhood change trend analysis by cellular automata to generate scenarios for more accurate and holistic simulations in urban basins. These scenarios predict the possibility of changes in projected volumes due to climate change in the recurrence time between events, as well as changes in the runoff pattern due to changes in Manning coefficients related to changes in cities. This research seeks to contribute scientific references for water resource management and minimise the effects of natural disasters in a holistic and proactive view of urban vulnerabilities within the context of tracing their evolution through the area they occupy. 2. DATA AND METHODS 2.1. Study area The study area for this research is the São Pedro stream watershed, located in the urban area of the city of Uberlândia, Minas Gerais, Brazil. This municipality is located in the Triângulo Mineiro mesoregion, in the state of Minas Gerais, in the Brazilian Cerrado biome, at latitude 18°55’23”S and longitude 48°17’19”W (Fig. 1), at an average altitude of 863 m. The climate of Uberlândia is typically hot and humid tropical, classified as Aw in the Köppen classification, with dry winters and rainy summers, and an average temperature of 22 °C [22]. The terrain is gently undulating in the northeast, varying to steep slopes in the southwest part of the basin. The catchment area of the São Pedro stream is almost entirely urbanised, with an area of approximately 49.0 km², and consists of three streams, the main one being the São Pedro stream, which receives the Jataí (1) and Lagoinha (3) streams, and two sub-basins, the Central and Sabiá Park. This basin is notable as it has a dam built within the Municipal Park, the main recreational park of the city. Built from concrete, the dam was designed to contain part of the surface runoff from the upstream areas, with a volume of 1,256 hm³ at an elevation of 856 m. Downstream, there are several residences, along with public areas such as supermarkets, shopping centres, football fields, swimming pools, the city hall and the municipal garden. [Insert Figure 1] 2.2. Methodological Procedures Factors such as the intensity and frequency of rainfall are the main causes of flooding, but factors linked to the scale and geometric resolution of the topography, river network, slope and soil sealing rate also have an influence on the assessment of flood vulnerability risks [23]. For this reason, the digital elevation model (DEM) was interpolated with a spatial resolution of 5 metres from contour lines. For land use and land cover (LULC), three images from the Landsat 7 and Landsat 8 satellites were used, for the years 2015, 2020 and 2025. Pan-sharpening was carried out with the panchromatic band to improve the geometric resolution of the multispectral images to 15m. These images were then classified into vegetation cover (including grass and tree covers), water bodies, exposed soil and sealed area, which combines buildings, roads and car parks. The historical rainfall series was obtained from the National Water and Basic Sanitation Agency (ANA) [24], a Brazilian agency that collates rainfall and river flow data from various sources. The station used for this research was 1948006, with coordinates 18º 59’ 18‘ and 48º 11’ 25’W, which provides daily data for the period from January 2009 to April 2025. 2.2.2. Change estimates and hydrological modelling The experimental plan for this research considered the evolution of the dynamics of use and occupation of the São Pedro stream basin area and the prediction of changes in the modelled rainfall pattern based on the recorded historical series. Figure 2 shows the research pipeline. [Insert Figure 2] 2.2.2.1. Changes in LULC Land use and land cover (LULC) changes were modelled using satellite images previously classified by use of the supervised minimum distance method, according to the classes mentioned. In addition to these, distances from roads, distance to water bodies and slope determined by buffers in the QGis software were considered. MLP models are commonly used to estimate land use changes, as the Land Change Modeler (LCM) module of the Terrset software [18] being used in several studies. However, in agreement with other authors [25, 26, 27, 28], MPL models function as a fully connected network, where each neuron in the layer is connected to all others in the next layer. Therefore, to process the images, it is necessary to transform them into a long vector, which causes the loss of topological relationships between pixels and their neighbours. Another problem arises from this, which is the increased chance of overfitting due to excessive complexity [29], making models of this type less robust and biased. On the other hand, convolutional networks (CNN) work on matrix structures, typical of images, and convolutional filters that allow spatial patterns to be captured, enriching the analyses and making the networks lighter when compared to MLPs [29]. In this research, U-net was selected because it is suitable for working with rasterised images. Another change proposed in this study applies to the stochastic estimation of class change by the Markov chain. The Markov chain estimates the transition potential of each pixel in the images to another LULC class using Cramer’s value (V), which is a parameter derived from the chi-square test and serves to assess the degree of adherence of the variables to each class. The organisation of these values in a matrix allows the quantification of the probabilities of change for each class in relation to those at each analysis interval. V values can vary between 0 and 1 (equation 1) but typical values are in the range of 0.40 to 0.80 [18]. V= (x2 / n.(q-1))1/² (1) Where: x² is the chi-square coefficient obtained from the frequencies of the classes, n is the sample size, and q is the smallest value in the rows and columns of the land use image. The classical approach considers the isolated pixel without taking into account the probabilities of changes in neighbouring pixels. Therefore, cellular automata (CA) is a promising approach to capture these changes in trends within the neighbourhood [29, 30] by creating CA-Markov. Combined with CNNs, CA-Markov chain allows the incorporation of change states into a grid of cells with variable values from the Markov matrix for each step. As a way to circumvent data stationarity problems, a typical situation in non-linear systems with variable amplitudes, [31, 32] recommend the use of Monte Carlo uncertainty assessment. This technique extends the original network dropout, allowing neural networks to quantify prediction uncertainty, as it captures random uncertainty caused by noise inherent in the data and epistemic uncertainty present in the model parameters [32]. Monte Carlo dropout was applied between layer connections in each network training interaction with 32-unit filters and a 20% rate, as a way to circumvent the problem of overfitting (Figure 3). [Insert Figure 3] 2.2.2.3. Pluviometry forecast For the precipitation simulation, an LSTM (Long Short-Term Memory) network was created, based on [19, 20, 21, 33, 34]. This type of neural network is specially designed to understand the temporal and dynamic sequences in serial data in which the output data is fed back at some point in the process before being processed again to obtain the final output [35, 36]. The advantage of this model is that the dependence on previous data, typical in series, allows the model to iteratively learn the dynamics of sequential data in temporal steps [35]. The rainfall station data contains daily records of 25 years of rainfall, temperature, and humidity collection. To create the LSTM network, a Python script was developed to analyse the data, model the network, and predict changes. The data were decomposed in terms of seasonality, trend and an irregular component. The trend was obtained by the moving average to avoid excessive smoothing by the kernel and mask localised rainfall peaks. Seasonality used the multiplicative form instead of the additive form, as it is considered that the amplitude of seasonality may vary over the years. The irregular component was calculated as the difference between the rainfall data recorded at the station and the value of the trend line and the seasonal component. These parameters were used in the normalisation of the series and training of the network. In this proposal, we use an attention model mechanism. The attention model (equation 2) used was dot-product self-attention, which learns which moments in the LSTM output time sequence are most important and highlights them via softmax before the final sum. A ReduceSumLayer function condenses this output into a single vector, which represents the entire sequence, taking into account which moments in time were most relevant to the prediction task. This approach allows the LSTM to learn to give more weight to the most relevant intervals within the decomposition window. Attention (Q, K, V) =softmax(QK/√(d_k ))∙ V (2) where Q, K, V ∈ R4×d are learned projections of Z. The attention output is mean-pooled to produce a spatially-aware embedding vector [37]. In this way, the characteristics of seasonality, irregularities, and trends are analysed in greater detail. This is achieved by calculating a score for each time interval based on the degree of adherence of the data to the final state of the LSTM network. This ensures that the network does not eliminate crucial data to the detriment of other less relevant data in the modelling. The LSTM was compiled with the attention model and a dropout rate of 20% to avoid overfitting effects. The Adam optimiser and the mean squared error loss function were used to train the model on the training data and evaluate its performance on the test data using RMSE and MAE. Validation was performed by comparing the modelled data with data from the last year of the historical series (April 2024 to April 2025), varying the decomposition time of the series and the data window (lookback parameter) for the forecast. Based on the network forecast values for the following months, the highest volumes (series peaks) were recorded and the highest from these values was taken as the possible extreme event. 2.2.2.4. Change in surface runoff Through the results simulated in the preceding steps, surface runoff for each scenario was simulated using the River Analysis System (HEC-RAS). The flood hydrographs were obtained using a peak value predicted by LSTM in each reference year, and the Huff method was applied to the 1st quartile. Using the hydrographs and the physical characteristics of the catchment area, such as the Curve Number (CN) and the soil impermeability rate identified by cross-referencing the geotechnical data collected in the region by [38] with the soil hydrological groups classified by [39], the evolution of runoff was determined according to the land use and occupation forecasts made for each year. These data were used to determine the evolution of surface runoff according to the LULC simulations. In order to calculate these hydrographs, the permeable portion of the land was considered using the SCS (Soil Conservation Service) method. This method separates the part of the runoff that infiltrates the soil from the part that drains away and assumes that the runoff in the impermeable portion of the basin is linked to the hydrographic system of the basin up to the outlet, where it is a function of the CN and the time of concentration. Based on the CN value, the time instant (∆t) in minutes is related to the time of concentration of the catchment in minutes to obtain the flow Q (t) (in m³.s-1) in the outlet of the São Pedro catchment. To simulate the flow in the main tributaries of the basin in the HEC-RAS model, additional hydrograph, geometric and hydraulic information was required as derived from the Digital Elevation Model (DEM), as well as the Manning coefficient according to the LULC in the simulated year, along with the adoption of permanent flow. Hydrographs were generated for the 2 and 6 hour rainfall durations in the HEC-HMS programme. Based on this procedure, it was possible to assess the impact of changes in LULC on surface runoff over the years 2025, 2030, 2035 and 2040. 3. RESULTS 3.1. Changes in LULC Figure 4 shows the LULC classification for the years 2015, 2020, and 2025, where urban expansion is observed, especially on the east and north sides of the São Pedro stream basin. In the east and northeast areas, there was a new wave of occupation due to the conversion of public to private land for new settlements. Over this ten-year period, vegetation areas were reduced by 26.2%, exposed soils by 37.5% and impervious areas increased by 39.9%, respectively 339.4 ha, 510.8 ha and 874.9 ha. The kappa indices for the LULC classification were all above 89.34% [Insert Figure 4] The probabilities of transition from one LULC class to another are presented by Markov values. Table 1 shows the values obtained for the transition from 2025 to 2035 between common Markov and CA-Markov. The Aggregating of the surrounding pixels led to different values for the classes Vegetation, for example, goes from a probability of 0.700 of remaining in the same class to 0.733, while the exposed soil class decreased from 0.342 to 0.290, as did other classes that had their probability of change values reduced in relation to the common Markov value. These results suggest that Cramer’s values (V) are overestimated, as these consider the statistical distribution without evaluating the context of neighbouring classes present in the environment of the analysed pixel. As a result, changes are expected to be slower over time when considered in the approach of this research. For each transition year, a new matrix was calculated in relation to the situation in the reference year. Thus, pixel transitions between classes were evaluated at fixed intervals of five years, but updated in relation to the estimates for the previous period. [Insert Table 1] Input image segmentation was performed using a deep U-Net, with 32×32 pixel patches, at a ratio of 1000 for training and 200 for validation, sampled using a relevance mask. The encoder comprises four convolutional blocks (32–512 filters) and a bridge of 1024 filters, mirrored by a decoder with transposed convolutions. Monte Carlo dropout was applied for uncertainty estimation with a value of 0.20, and for training, the Adam method was used with a learning rate of 0.001, batch size 16 and up to 20 epochs, producing pixel-by-pixel segmentation maps suitable for complex spatial analyses, such as flood risk assessment in urban basins. For comparison, the values in Table 1 were evaluated by an MLP network implemented in Clark Laboratories’ Land Change Modeler (LCM-TerrSet) software [23] and evaluated by the U-net proposal with CA-Markov. While in the MLP, the overall average accuracy index of the network reached 70.72%, the combined approach of this research achieved 83.85%, kappa index 0.7723 and F1-Score 0.844. Figure 5 shows the transition potentials by region of the basin representation for 2030, with the highest values for the vegetation classes for exposed soil and exposed soil to impervious area, with a strong trend in the eastern and southern areas. [Insert Figure 5] It worth mentions that in highly urbanised basins such as São Pedro, the drivers of change (such as motorways, urbanised areas and urban restrictions) tend not to undergo major changes over the years and, for this reason, the main differences are more evident in the opening up of new areas rather than in the densification of sparse areas between urban conglomerates, for example. Based on these changes, the characteristics of the catchment were extracted using GIS and obtained a length of 11401.19m, a slope of 0.013972 m/m, an upstream elevation of 944.98m and a downstream elevation of 785.68m. In order to quantify a Manning coefficient for each year, the stream was divided into sections of the sub-basins (Parque do Sabiá, Jataí, São Pedro, Central and Lagoinha) and a weighted average was used that considered the distances with the Manning average, which obtained the values for each year as shown in Table 2: [Insert Table 2] Table 3 shows the area values, and the results of average CN and average impermeability for the years studied and for each sub-basin. [Insert Table3] 3.2. Rainfall forecast from LSTM One of the most important characteristics of LSTM is determining the time window for analysing forecasts (lookback), as small values can lead to inadequate scaling and large values can make it impossible to capture the pattern properly, tending towards the average of the series values. In addition, there is an inverse relationship between the size of the series decomposition window used in network training and the lookback window for forecasting. After successive configurations, we arrived at a value of 90 days and 120 days for the decomposition window, which presented an RMSE value of 0.1087, the lowest value tested. The LSTM layer was performed with 50 units (return sequences) followed by an attention mechanism to weight relevant periods of the series (ReduceSum) and a dropout rate of 0.2 for uncertainty estimation. To avoid negative predictions, a ReLU layer was used. The model was trained with the Adam method, batch size of 32 with 50 epochs and a validation ratio of 20%. The Monte Carlo method was again used for 95% confidence intervals, providing robust point estimates and uncertainty limits for the network’s predictions. Figure 5 shows the historical series and its decomposition into seasonality, trend, and an irregular component. The series shows a historical peak of 111 mm on 12/09/2020 and alternating months of rain in well-defined seasons. However, the linear trend line has pointed to a decrease in values at a rate of 0.011 mm/year, while the temperature has tended to increase by 0.07 °C/year. As a result, the trend line shows a stable slope until 2015 and a downward trend in rainfall since then. These differences are reflected in seasonality with a slight decline in accumulated rainfall peaks in recent years and an irregular component with negative peak values (Fig. 6). [Insert Figure 6] Based on this decomposition, the LSTM network was trained and validated with data from the series itself, as shown in Figure 6. Validation was performed by comparing the period from April 2023 to April 2024 with the forecast results from the network. During this period, the two highest peaks recorded were 95.1 mm and 63.2 mm, and the test validations simulated peaks of 88.03 mm and 63.12 mm for dates that are close together (Figure 7). Although future forecasts are below these values, reaching 58.45 mm, they may be influenced by components from the previous year. This indicates the need for constant calibration with updated data. The results of the LSTM+AM forecast were compared with the ARIMAX values. The AutoRegressive Integrated Moving Average with eXogenous regressors (ARIMAX) model extends the original model by performing a multivariate analysis through the inclusion of complementary variables such as temperature and pressure in the rainfall forecast analysis. According to Rizvi, Sahu and Rana [40] and Ponce et al. [41], the inclusion of external variables improves the explanatory power of the model while preserving the statistical interpretability of the results. However, when comparing the metrics of LSTM and Arimax, RSME values of 10.0675 and MAE values of 3.7765 were found for LSTM, while the ARIMAX method presented values of 8.8392 and 6.3153, respectively. In Arimax the peaks were reduce tending to the mean of the series was could be interpreted as difficult to modelling the seasonality as expected. [Insert Figure 7] Based on these results, hyetographs were simulated using the Santa Bárbara method and flood spots were simulated based on LULC changes. Considering the maximum peak of LSTM with a value of 84.56 mm as an event for simulation. 3.3. Hyetograms and changes in flows By performing the simulations as described above, that is combining the LSTM network forecast and the LULC changes, it was possible to generate the pluviograms for 2-hour and 6-hour rainfalls that generate direct surface runoff into the basin. These discretised rainfalls were entered into HEC-RAS along with the CN and sealing rate values from Table 3. Figure 8 shows the pluviograms of the 2-hour and 6-hour rainfall generated using the Huff 1st quartile method. These rainfalls are responsible for generating the direct surface runoff flows into the sub-basins. One notes that between these rainfalls there is a difference in the highest value of 23.66 mm for the 2-hour rainfall and 12.71 mm for the 6-hour rainfall. This latter value is more distributed over the period in which it occurs, with a considerable reduction from 180 minutes onwards. [Insert Figure 8] These discretised rainfalls, the area of each sub-basin, the average NPP values and the average soil sealing rates for the four periods studied (2025, 2030, 2035 and 2040) were entered into the HEC-RAS programme, which generated four hydrographs at the outlet of the basin, shown in Figures 9 and 10. Comparative analysis of these hydrographs reveals gradual increases in peak flow over time as LULC and rainfall changes in the basin. [Insert Figure 9] [Insert Figure 10] Table 4 summarise these peaks variation when compared to the current situation (2025). For example, the 2-hour rainfall presents an increase of 697.0 m³/s in 2025 growing up to 722.6 m³/s in 2040 (+3.67%), while the 6-hour rainfall increases from 374.9 m³/s to 392.7 m³/s (+4.75%). One notes here a greater increase in the first variation and smaller increases in the following years. [Insert Table 4] While these percentage increases may appear modest, it is essential to consider them in the context of model uncertainty. Using Monte Carlo Dropout from the LSTM forecasts and propagating LULC scenario variability, the 95% confidence intervals for the projected peak flows indicate that the observed changes are generally larger than the expected stochastic variability, suggesting a statistically meaningful trend in peak runoff. The sensitivity analysis also shows that changes in the rate of soil sealing are the main determinants of these changes, while variations in rainfall intensity have a relatively minor effect on peak flows. In a practical perspective, a 4.75% increase corresponds to an estimated rise of approximately 1.9 m in the flood level at the basin outlet, resulting in an expansion of the flooded area by about 62.9 hectares and exposing populations currently in safe areas. These results are explained by the topographical characteristics of the basin, which is composed of plateaus interspersed with deeper valleys where the basin’s channels are located. Finally, this shows that the hydraulic infrastructure designed based on current land cover may become undersized in the next two decades if such changes are not incorporated into planning. 4. CONCLUSIONS This study presented an innovative approach to integrating anthropogenic dynamics and the effects of climate change on flooding in urban basins, demonstrating high predictive potential and practical applicability. The combination of historical series of land use, precipitation, temperature, and humidity with advanced machine learning techniques made it possible to construct realistic scenarios capable of anticipating extreme events and delimiting risk areas. This approach represents a significant advance in urban hydrological modelling, especially by incorporating human and climatic variables into the same analytical framework to model the complex rainfall-runoff relationship in urban catchments. In the case of the São Pedro Stream basin, it was found that short-duration events (<2 h) generate significant volumes of surface runoff, reinforcing the role of increasing impermeability in urban vulnerability. From a computational point of view, attention-enhanced LSTM networks outperformed conventional LSTM models and models such as ARIMAX, demonstrating superior predictive power for multivariate data. Similarly, the use of surrounding pixels proved to be more realistic than the conventional technique of point-by-point land use change analysis. Combined, the results highlight the need for integrated management of this upstream basin, and this approach enables real-time forecasting, supports decision-making for urban planning, and provides actionable insights for infrastructure protection. Despite the results, we identified some limitations, mainly related to the sensitivity of LULC predictions to image classification quality and sensor spatial resolution. The CA-Markov model assumes homogeneous transitions in space, which may not reflect complex urban dynamics, and the HEC-RAS simulation considers static infrastructure conditions, not incorporating future changes such as drainage works or urban expansion. From a computational point of view, the incorporation of attention mechanisms in neural networks proved to be superior to the isolated use of LSTM with multivariate data but still requires continuous calibration of the network hyperparameters for real-time monitoring. The approach proposed in this article paves the way for the integration of AI systems and hydrological simulations to support the creation of real-time forecasting tools, which are particularly important for urban centres with a history of extreme events, as well as supporting urban and environmental zoning decisions. DECLARATION OF COMPETING INTEREST The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. ACKNOWLEDGEMENTS The authors would like to thank the Research and Development Program, P&D ANEEL, along with EMAE (grant number 00393-0015/2022). DATA AVAILABILITY Available on GitHub: available after approval REFERENCES [1] Leng, Z.; Chen, L.; Yang, B.; Li, S.; Yi, B. (2024). An extreme forecast index-driven runoff prediction approach using stacking ensemble learning. Geomatics, Natural Hazards and Risk , 15. https://doi.org/10.1080/19475705.2024.2353144 [2] CRED, Centre for Research on the Epidemiology of Disasters. (2023). Disasters in Numbers. Brussels: CRED. https://files.emdat.be/reports/2023_EMDAT_report.pdf. 2023 (acessed 17th October 2024). [3] Al-Fugara, A.; Mabdeh, A.N.; Alayyash, S.; Khasawneh, A. (2023). Hydrological and Hydrodynamic Modeling for Flash Flood and Embankment Dam Break Scenario: Hazard Mapping of Extreme Storm Events. Sustainability , 15, 1758. https://doi.org/10.3390/su15031758 [4] Gaagai, A.; Aouissi, H.A.; Krauklis, A.E.; Burlakovs, J.; Athamena, A.; Zekker, I.; Boudoukha, A.; Benaabidate, L.; Chenchouni, H. (2022). Modeling and Risk Analysis of Dam-Break Flooding in a Semi-Arid Montane Watershed: A Case Study of the Yabous Dam, Northeastern Algeria. Water , 14, 767. https://doi.org/10.3390/w14050767 [5] Zhang, K; Mei, C.; Liu, J.; Wang, J.; Song, T.; Shi, H.; Dong, L.; Li, Y.;Zhang, A. 2025. Impact assessment of urban waterlogging on roads trafficability and emergency sites accessibility under extreme rainfall events based on numerical modelling . International Journal of Disaster Risk Reduction , 119. https://doi.org/10.1016/j.ijdrr.2025.105285 [6] Maranzoni, A.& Tomirotti, M. (2023). Three-Dimensional Numerical Modelling of Real-Field Dam-Break Flows: Review and Recent Advances. Water , 15, 3130. https://doi.org/10.3390/w15173130 [7] Mattas, C.; Karpouzos, D.; Georgiou, P.; Tsapanos, T. (2023). Two-Dimensional Modelling for Dam Break Analysis and Flood Hazard Mapping: A Case Study of Papadia Dam, Northern Greece. Water , 15, 994. https://doi.org/10.3390/ w15050994 [8] Marangoz, H. O. & Anilan, T. (2022). Two-dimensional modeling of flood wave propagation in residential areas after a dam break with application of diffusive and dynamic wave approaches. Natural Hazards . 110, pp 429–449. https://doi.org/10.1007/s11069-021-04953-w [9] Navarra, C.; Kucher, K.; Neset, T.; Villaro, C. G.; Schück, F.; Unger, J.;Vrotsou, K. (2025). (2025). Leveraging visual analytics of Volunteered Geographic Information to support impact-based weather warning systems. International Journal of Disaster Risk Reduction. 126 . https://doi.org/10.1016/j.ijdrr.2025.105562 [10] Tanaka, K; Akaishi, K.; Yokota, T. 2025. Public preferences for flood warning improvements: An inferred valuation approach addressing social desirability bias. International Journal of Disaster Risk Reduction , 126. https://doi.org/10.1016/j.ijdrr.2025.105571 [11] Moradian,S.;Gharbia,S.; Haghighi, A. T. & Olbert, I. A. (2025). Modelling extreme precipitation projections under the effects of climate change: case study of the Caspian Sea. International Journal of Water Resources Development , 41,pp 57-77. https://doi.org/10.1080/07900627.2024.2400505 [12] Guo, K.; Guan, M.; Yu, D.: (2021). Urban surface water flood modelling – a comprehensive review of current models and future challenges. Hydrology and Earth System Sciences , 25, pp 2843–2860. https://doi.org/10.5194/hess-25-2843-2021. [13] Bolick, M.M.; Post, C.J.; Naser, M.Z.; Forghanparast, F.; Mikhailova, E.A. (2023). Evaluating Urban Stream Flooding with Machine Learning, LiDAR, and 3D Modeling. Water , 15, 2581. https://doi.org/10.3390/w15142581 [14] Liu, B.; Tan, X.; Gan, T. Y.; Chen, X.; Lin, K.; Lu, M.; Liu, Z. (2020). Global atmospheric moisture transport associated with precipitation extremes: Mechanisms and climate change impacts. Wiley Interdisciplinary Reviews: Water . 7 https://doi.org/10.1002/wat2.1412. [15] Huang, S.; Chang, J.; Huang, Q.; Chen, Y. (2015). Identification of abrupt changes of the relationship between rainfall and runoff in the Wei River Basin, China. Theoric Application Climatology . 120, 299–310. https://doi.org/10.1007/s00704-014-1170-7 [16] Cea, L.; Costabile, P. (2022). Flood Risk in Urban Areas: Modelling, Management and Adaptation to Climate Change: A Review. Hydrology , 9, 50. https://doi.org/10.3390/hydrology9030050 [17] Fathi, M.M., Al Mehedi, M.A., Smith, V. et al. 2025. Evaluation of LSTM vs. conceptual models for hourly rainfall runoff simulations with varied training period lengths. Scientific Reports , 15, 15820. https://doi.org/10.1038/s41598-025-96577-4 [18] Eastman, J. (2016). TerrSet - Geospatial Monitoring and Modeling System: Manual. Worcester, Massachusetts, USA, Clark University, 1-342. [19] Mhedhbi, R. & Erechtchoukova, M.G. (2025). Assessing the Impact of Rainfall Nowcasts on an EncoderDecoder LSTM Model for Short-Term Flash Flood Prediction. Water Resources Management . 39, pp1623–1638. https://doi.org/10.1007/s11269-024-04037-x [20] Pan, X.; Hou, J.; Gao, X.; Chen, G.; Li, D.; Imran, M.; Li, X.; Yang, N.; Ma, M.; Zhou, X. (2025). LSTM Model-Based Rapid Prediction Method of Urban Inundation with Rainfall Time Series. Water Resources Management . 39:661–688. https://doi.org/10.1007/s11269-024-03972-z [21] Coutinho, E. R.; Madeira, J.G.F.; Borges, D.G.F.; Springer, M.V.; de Oliveira, E.M.; Coutinho, A.L.G.A. (2025). Multi-Step Forecasting of Meteorological Time Series Using CNN-LSTM with Decomposition Methods. Water Resources Management . https://doi.org/10.1007/s11269-025-04102-z [22] Silva, V. J. da; Silva, C. R. da; Almeida, L. da S.; Carvalho, H. P; Camargo, R de. (2018). Mobile transect for identification of intra-urban heat islands in Uberlandia, Brazil. Ambiente & Água - An Interdisciplinary Journal of Applied Science. ISSN 1980-993X. https://doi.org/10.4136/ambi-agua.2187 [23] Wu, J.; Chen, X.; Lu, J. (2022). Assessment of long and short-term flood risk using the multi-criteria analysis model with the AHP-Entropy method in Poyang Lake basin. International Journal of Disaster Risk Reduction, 75. https://doi.org/10.1016/j.ijdrr.2022.102968 [24] Agência Nacional de Águas (ANA). Hidroweb: Sistemas de informações hidrológicas . http://hidroweb.ana.gov.br/. (acessed 4th May 2025). [25] Parracciani, C., Gigante, D., Mutanga, O., Bonafoni, S.; Vizzari, M. (2024). Land cover changes in grassland landscapes: combining enhanced Landsat data composition, LandTrendr, and machine learning classification in google earth engine with MLP-ANN scenario forecasting. GIScience & Remote Sensing , 61(1). https://doi.org/10.1080/15481603.2024.2302221 [26] Alqadhi, S., Mallick, J., Balha, A. et al. (2021). Spatial and decadal prediction of land use/land cover using multi-layer perceptron-neural network (MLP-NN) algorithm for a semi-arid region of Asir, Saudi Arabia. Earth Science Information . 14, 1547–1562. https://doi.org/10.1007/s12145-021-00633-2 [27] Sabri, A., Bahi, H., Bounoua, L.; Tahiri, M.; Tweed, S.; LeBlanc, M.; Bouramtane, T.; Malah, A.; Kacimi, I. (2024). Predictive modelling on Spatial–temporal Land Use and Land Cover changes at the Casablanca-Settat Region in Morocco. Modelling Earth System Environment . 10, pp 6691–6714. https://doi.org/10.1007/s40808-024-02107-y [28] Devi, A. R.; Shimrah, T. (2023). Modeling LULC using Multi Layer Perceptron Markov change (MLP MC) and identifying local drivers of LULC in hilly district of Manipur, India. Environmental Science and Pollution Research, 30:68450–68466. https://doi.org/10.1007/s11356-023-27153-4 [29] Driss, S.; Soua, M.; Kachouri, R.; Akil, M. (2017). A comparison study between MLP and Convolutional Neural Network models for character recognition. Environmental Science and Pollution Research. 30. https://doi.org/10.1117/12.2262589. [30] Beyene, E.; Minale, A.S. (2023). Modeling urban land use dynamics using Markov-chain and cellular automata in Gondar City, Northwest Ethiopia. Chinese Journal of Population, Resources and Environment , 21(2), pp 111-120, https://doi.org/10.1016/j.cjpre.2023.06.007. [31] Van Katwyk, P.; Fox-Kemper, B.; Seroussi, H.; Nowicki, S.; Bergen, K. (2023). A Variational LSTM Emulator of Sea Level Contribution From the Antarctic Ice Sheet. Journal of Advances in Modeling Earth Systems . 15. https://doi.org/10.1029/2023MS003899. [32] Kummaraka, U., & Srisuradetchai, P. (2025). Monte Carlo Dropout Neural Networks for Forecasting Sinusoidal Time Series: Performance Evaluation and Uncertainty Quantification. Applied Sciences , 15(8), 4363. https://doi.org/10.3390/app15084363 [33] Arabaci, Hayri & Ucar, Kursad & Çimen, Halil. (2024); Examining the influence of sampling frequency on state-of-charge estimation accuracy using long short-term memory models. Electrical Engineering . 106, 6449-6462. https://doi.org/10.1007/s00202-024-02392-x. [34] Xiang, Z., Yan, J., & Demir, I. (2020). A rainfall‐runoff model with LSTM‐based sequence‐to‐sequence learning. Water Resources Research , 56. https://doi.org/10.1029/2019WR025326 [35] Ouma, Y. O. ; Cheruyot, R.; Wachera, A. N. (2022). Rainfall and runoff time‑series trend analysis using LSTM recurrent neural network and wavelet neural network with satellite‑based meteorological data: case study of Nzoia hydrologic basin . Complex & Intelligent Systems . 8, 213–236. https://doi.org/10.1007/s40747-021-00365-2 [36] Gao, P.; Du, W.; Lei, Q.; Li, J.; Zhang, S.; Li, N. (2023). NDVI Forecasting Model Based on the Combination of Time Series Decomposition and CNN – LSTM. Water Resources Management . 37, 1481–1497. https://doi.org/10.1007/s11269-022-03419-3 [37] Adibfar, A.; Davani, H. 2025. PrecipNet: A transformer-based downscaling framework for improved precipitation prediction in San Diego County. Journal of Hydrology: Regional Studies , 62. ISSN 2214-5818. https://doi.org/10.1016/j.ejrh.2025.102738 [38] Andrade, F. de. (2005). Mapeamento Geotécnico Preliminar em Escala de Semi- detalhada (1:25.000) da Área de Expansão Urbana de Uberlândia . [Master Dissertation, Universidade Federal de Uberlândia (UFU)]. pp 135. https://repositorio.ufu.br/handle/123456789/14105 [39] Sartori, a.; Lombardi Neto, F.; Genovez, A. M. (2005). Classificação Hidrológica de Solos Brasileiros para a Estimativa da Chuva Excedente com o Método do Serviço de Conservação do Solo dos Estados Unidos. Parte 1: Classificação . Revista Brasileira de Recursos Hídricos. 10, 4. https://doi.org/10.21168/rbrh.v10n4.p5-18 [40] Rizvi, M. F.; Sahu, S.; Rana, S. (2024). ARIMA model time series forecasting. International Journal for Research in Applied Science & Engineering Technology (IJRASET), 12(5), 3782–3786. https://doi.org/10.22214/ijraset.2024.62416 [41] Ponce, C. M.; Caballero, A.; Huaman, C.; Gamarra, D.; Rojas, A. (2025). Analysis of time series forecasting models applied to health and environmental data: SARIMAX approach in Latin America. BMC Public Health , 25(1), 24125. https://doi.org/10.1186/s12889-025-24125-w Supplementary Material File (hydrological_processes_synergisticeffects_tables.docx) Download 18.78 KB Information & Authors Information Version history V1 Version 1 31 October 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords climate change disaster prevention flash flooding simulation lstm- attention models modelling rainfall-runoff estimation Authors Affiliations Marcio Schmidt 0000-0003-2716-2360 [email protected] Universidade Federal de Uberlandia View all articles by this author Carlos Eugênio Pereira Universidade Federal de Uberlandia View all articles by this author Alan Petrônio Pinheiro Universidade Federal de Uberlandia View all articles by this author Metrics & Citations Metrics Article Usage 209 views 162 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Marcio Schmidt, Carlos Eugênio Pereira, Alan Petrônio Pinheiro. Synergistic Effects of Climate and LULC Change on Urban Rainfall–Runoff Dynamics: A Data-Driven Framework With Lstm–Attention Models. Authorea . 31 October 2025. DOI: https://doi.org/10.22541/au.176192481.11762995/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.176192481.11762995/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a00ba0017cb80704',t:'MTc3OTYxODYyNg=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();