Using machine learning and satellite data to improve flood forecasting: the case of the Ouémé basin at the Bétérou outlet

Water-related disasters, in general, and floods, in particular, are significant threats to societies and ecosystems worldwide. These events are responsible for considerable loss of life and economic damage, with more than 300,000 deaths and around 1.7 trillion dollars in financial losses recorded between 2001 and 2018 (Lee et al. 2020; WMO 2021). According to statistics from the United Nations Office for Disaster Risk Reduction (CRED 2015), some 150,061 floods have occurred worldwide, causing 157,000 deaths and 11.1% of global disaster victims between 1995 and 2015. Between 1970 and 2000, reports of medium- and large-scale disasters averaged around 90–100 per year, but between 2001 and 2020, the reported number of such events increased to 350–500 per year (UNDRR 2023).

Benin is heavily affected by cyclical flooding. These floods are caused by various natural, man-made, and organizational factors. In 2010, the total damage caused by floods to Benin's economy amounted to nearly USD 160 million (World Bank 2011). In 2021, flooding on the country's main rivers (Niger, Ouémé, and Mono Rivers) and their tributaries affected 38 communes, resulting in a great deal of damage. The Ouémé River, with its high agricultural potential, is the richest valley in Benin. A large number of economic activities (agriculture, fishing, livestock farming, and tourism) are developed in this area. Unfortunately, the population of this area is regularly hit by recurrent floods, which cause material and economic damage and loss of life (Quenum et al. 2022). The growing impact of floods is exacerbated by climate change, which is intensifying the frequency and severity of extreme weather events (Tabari 2020; Man et al. 2023). In this context, numerous studies have been carried out in the basin with the aim of predicting flows using traditional hydrological models. These include global conceptual models (Kodja et al. 2018; Bossa et al. 2024), physics-based global models (Alamou 2011; Biao et al. 2016; Biao 2017) and distributed physics models (Hector et al. 2018; Bodjrènou et al. 2023). Hydraulic models such as HEC–HMS (Houngue 2020) were also used. Although these models give satisfactory results for simulating flows (the Nash–Stucliffe coefficient is often between 0.7 and 0.85), they nevertheless have great difficulty in reproducing flood flows and low flows. Floods are often underestimated, which poses a serious problem for flood management. Faced with this challenge, the advent of modern technologies offers an unprecedented opportunity to improve flood prediction.

To address the problem of flooding, structural and non-structural methods were needed (Man et al. 2023). Structural methods are the visible flood control measures, such as the construction of dams, dykes, and weirs. At the same time, non-structural methods such as flood forecasting models and systems, which facilitate disaster preparedness planning, have played a major role. Computer science and hydrology development has boosted the efforts made in flood forecasting and flood management in recent decades. One of the most interesting methods introduced into predictive hydrology is artificial intelligence (AI).

Integrating AI into hydrological modeling is a significant advance, adding substantial value to the management and analysis of large quantities of data (Zhang et al. 2021; Jones et al. 2023). This integration enables models to improve themselves through continuous learning. The diversity of machine learning (ML) techniques, such as artificial neural metworks (ANN), support vector machines, multi-layer perceptrons, decision trees, adaptive neurofuzzy systems, wave neural networks, and probabilistic expert systems, enriches the field of possibilities for flood prediction and analysis (Nevo et al. 2022).

The improved ANN models have reduced the root mean square error (RMSE) by 10% for test datasets (synthetic events) and by 16% for real events, underlining the effectiveness of these networks in the hydrometeorological context (Abarghouei & Hosseini 2016; Khairuddin et al. 2019; Zhu et al. 2021). In terms of deep learning, recurrent neural networks (RNNs), including their variants such as gated recurrent units, are being evaluated for their ability to deliver near-reality predictions (Talei 2022), which is essential for hydrological forecasting and water resource management. Long short-term memory (LSTM) neural network models are being considered for flood vulnerability prediction and could improve the accuracy of runoff predictions and reduce model training costs and time (Hu et al. 2018; Bai et al. 2021; Yin et al. 2021).

A model combining the convolutional neural network (CNN) and LSTM demonstrated the best performance in predicting runoff peaks during several flooding events, with a Nash–Sutcliffe efficiency (NSE) index greater than 0.9 for five of them (Durrani et al. 2023; Zhou et al. 2023). Another study suggested an LSTM model for flood forecasting, using daily flow and precipitation data as input, under conditions where the characteristics of the datasets that can influence model performance were of particular interest (Le et al. 2019; Fang et al. 2021; Cho et al. 2022; Durrani et al. 2023).

Despite the AI model's prowess, the quality and quantity of the input data are a determining factor in the quality of the results. The challenge, in the case of the Ouémé catchment, is the availability of observed meteorological data. As in other African watersheds, the Ouémé is a watershed that suffers from a scarcity of data due to a lack of density in the measurement network and poor monitoring of this network. In these conditions, the possibility of adding satellite data to the model's input was explored. In fact, remote sensing and AI offer many possibilities for flow simulation (Roderick et al. 2021). Satellite remote sensing has radically changed hydrology by providing extensive and high-resolution data, which is crucial for analyzing hydrological phenomena (Crétaux et al. 2017; Huang et al. 2018; Kittel et al. 2018; Rodell & Reager 2023). Remote sensing has made possible the simulation of flows and other hydrological processes in poorly gauged or ungauged catchments (Kim et al. 2019; Nickles et al. 2020; Kittel et al. 2021; Lamine et al. 2021; Revel et al. 2021; Xiong et al. 2021). The integration of different data sources, including satellite and in situ, can be complex and requires advanced data processing and analysis methods (Samadi 2022). The practical application requires attention to the specific conditions of the study area, calling for ongoing research to refine these tools and maximize their applicability in the field (Zhou et al. 2023). Our study focuses on exploiting these technological advances to improve flood prediction in the Ouémé basin at Bétérou, an emblematic example of river systems in dry tropical environments where in situ data are insufficient. The aim of this study is threefold: (1) evaluate the use of satellite data and existing AI techniques for flood prediction, (2) develop a specific flood prediction model, and (3) map flooding risk areas. Our research aims not only to make a significant scientific contribution but also to provide practical tools for making communities more resilient to flooding.

A variety of data was collected for this study. Historical climate data, including precipitation, temperature, and wind speed, were obtained from the European Centre for Medium-Range Weather Forecasts via the ERA database, providing hourly records essential for understanding local atmospheric processes. These data were extracted in NetCDF format. River flows, measured by the Amma-Catch network, indicate the response of the catchment area to precipitation (AMMA-CATCH 1990). Finally, satellite images from the US Geological Survey (USGS) (EarthExplorer (usgs.gov)) have been used to characterise the topography of the basin and are associated with water height data obtained from the General Water Directorate of Benin to identify areas at risk of flooding.

The ConvLSTM model (Shi et al. 2015), an advanced architecture specially designed for the efficient processing of spatio-temporal data, was implemented. ConvLSTM is particularly suitable for analyzing spatio-temporal meteorological variables such as precipitation, temperature, and wind, and for predicting river discharge, a temporal datum (Shi et al. 2015; Xingjian et al. 2015). The first processing step in the ConvLSTM model consists of a convolutional layer, which applies convolution kernels to the input data to extract key spatial features. This approach effectively identifies local patterns and spatial interdependencies in weather data, as demonstrated in the work of Krizhevsky et al. (2012). The ability of CNN to process spatial data is well documented and essential to our analysis (LeCun et al. 1998). After the convolutional layer, the data are processed by LSTM layers, known for their effectiveness in capturing long-term temporal dependencies in sequential data. LSTM is therefore used to interpret the extracted spatial features, enabling the model to understand the temporal dynamics of the input variables and accurately predict river flow. This ability of LSTM to manage temporal sequences is highlighted in the work of Hochreiter & Schmidhuber (1997). The adopted model therefore takes advantage of the strengths of CNN for recognizing spatial patterns and LSTM for analysing temporal sequences. This synergy makes capturing both the spatial and temporal relationships in the data possible, which is crucial for predicting complex phenomena such as river flow. An LSTM model has been designed to analyze and predict temporal sequences to map areas at risk of flooding. In our context, this model is used to predict the height of water as a function of the discharge estimated by our flood forecasting model. The LSTM model was trained using historical water level and discharge data to identify the relationship between river discharge and corresponding water levels. Once optimized and trained, the LSTM model was used to predict water levels based on predicted flows. These forecasts, generated by the LSTM model, were then integrated into the Global Mapper software to carry out flood surge. Depending on the water level, this software allows the water to be simulated and the areas prone to flooding to be visualized.

CNN is a class of deep architectures in AI that is specialized in the analysis of visual data, mimicking the way the human brain processes visual information through its ability to automatically recognise features and patterns in images (Kareem et al. 2021). In CNN architectures, three main types of layers are distinguished by their specific functions. The first, the convolution layer (CONV), is the heart of the CNN. It uses filters to perform convolution operations on the input, capturing local features such as edges and textures. Key hyperparameters of this layer include filter size (F), stride (S), and number of filters, which determine the dimensionality and depth of the feature maps produced. The second key layer is the pooling layer (POOL), which performs a sub-sampling operation. This layer reduces the dimensionality of the feature maps and introduces spatial invariance into the network, which is essential for reducing computational complexity and avoiding overfitting. The final layer, the fully connected layer, interprets the features detected by the previous layers to perform tasks such as classification. In this layer, each neuron is connected to all neurons in the previous layer, allowing full integration of learned features (Chen et al. 2023). In addition to these layers, CNN includes important aspects such as filter hyperparameters, including filter dimensions, stride, and zero padding, which adjust the size of activation maps. Activation functions such as Rectified Linear Unity and Softmax are also crucial. The former introduces non-linearities into the network, and the latter transforms output scores into probabilities in classification tasks. Every element of a CNN, from layers to hyperparameters, plays a vital role, making these networks particularly effective for applications such as image recognition and classification.

LSTM is an RNN architecture widely used in deep learning. In sequence prediction challenges, LSTM networks are a type of recurrent neural network that can learn order dependency. The output of the previous step is used as input to the current step in RNN. By default, the LSTM can retain and memorize information for both short and long times. A conventional LSTM unit comprises a cell, an input gate, an output gate, and a forget gate. These three gates control the flow of information into and out of the cell, and the cell remembers values over arbitrary time intervals (Picornell et al. 2023).

The operation of the gates in LSTM is as follows:

• – Forgetting gate: Determines which information from the previous state (⁠⁠) and the current input (⁠⁠) should be forgotten using a sigmoid function.

• – Input gate: Selects the new information to be retained using a combination of the sigmoid and hyperbolic tangent functions.

• – Output gate: Decides what information to pass to the hidden state using the current state of the cell (C_t) and a sigmoid function. The following equations can describe the operations performed by an LSTM:

  

  (1)

  

  (2)

  

  (3)

  

  (4)

  

  (5)

  

  (6)

  where represents a sigmoid function, W denotes a weight matrix, b indicates a bias, and ⊙ means the Hadamard product operation (Cho et al. 2022).

The water level rise and flood simulation tool was used to model flood risk areas for different water levels. This tool models water cover by raising the water level to a specified depth from a selected surface feature, such as a floodplain. It identifies all points in the terrain upstream of the selected features whose entry into these areas is less than the specified water level elevation using algorithms similar to those used in catchment delineation. This makes it possible to accurately determine the areas that will be flooded. The tool also takes into account topographical features likely to prevent or restrict flow, such as dykes, buildings, and other natural obstacles. This ensures that the simulation is realistic and avoids inaccurate results, such as the filling in of depressions that cannot be reached by water.

In this work, the metrics used to evaluate the performance of simulations are mainly the coefficient of determination (R²), the RMSE, and the Nash criterion.

An RMSE close to zero indicates better model accuracy.

The NSE varies from to 1, and values close to 1 indicate a better reproduction of flows by the model.

Daily flows at the Bétérou outlet vary from 0 to 621 m³/s (Figure 3(e)), with a mean of 54.71 m³/s and a standard deviation of 100.70 m³/s. This indicates a wide dispersion of observed flow around the mean.

In this work, we used Bayesian optimization from the Scikit-Optimize library (skopt) to explore the space of hyperparameters and minimize the loss function while taking into account previous evaluations of the loss function to guide the search.

Figure 4(b) shows the hyperparameters obtained by the LSTM model optimisation process. Each graph shows how the interaction between two different hyperparameters influences the overall performance of the model, measured in terms of accuracy during validation. Regions in light shading correspond to better performance, indicated by red stars marking optimal configurations.

Graphs on the diagonal represent performance as a function of a single hyperparameter, while those off the diagonal reveal the combined effect of two hyperparameters. These visualisations provide an immediate overview of significant trends and interactions between hyperparameters. For example, it can be seen that the learning rate and the number of LSTM units are intimately linked to optimal model performance, highlighting their importance in fine-tuning the model. The red stars indicate the configurations that gave the best performance.

The optimal configuration derived from this analysis is used for further experimentation. This configuration, which maximises validation accuracy, suggests a promising ability of the model to generalise beyond the training dataset.

Graphs (b) and (c) of Figure 5 show the performance of the LSTM model in calibration and validation. It can be seen that the LSTM model is well suited to handling sequential data. The Nash and the coefficient of determination of the simulated water heights are equal to 0.95 while the RMSE is 0.046. More importantly, the model is able to simulate peaks well, which is important for our analysis, which focuses on extreme values.

According to Dehghani et al. (2023), neural networks (LSTM) are particularly effective in contexts where the spatial distribution of data is a key factor. This efficiency is notable, for example, in small river basins equipped with well-distributed rainfall stations. This characteristic explains the high performance of our LSTM model in similar applications. In addition, Smith et al. (2020) implemented LSTM models to predict water flows in the context of catchment modeling, obtaining a coefficient of determination (R²) of 0.85. This result is lower than for our ConvLSTM model, which highlights the importance of taking into account the spatial properties of the phenomenon studied, which improves the accuracy of predictions. In another study, Johnson et al. (2019) compared various ML techniques for forecasting water flows. The performance of their LSTM model, with a coefficient of determination of 0.96, as well as that of our own model, confirms the right decision to use these approaches to predict water levels. Chen et al. (2019) used a ConvLSTM network to model rainfall runoff, achieving good results in terms of water flow predictions. Durrani et al. (2023) also have the same results using ConvLSTM and LSTM. Xingjian et al. (2015) have developed an ML approach for nowcasting precipitation, similar to our model for flow prediction. According to their work, ConvLSTM is particularly effective at better capturing spatio-temporal correlations in data. There are two main reasons for this. First, ConvLSTM is well suited to handling boundary conditions, given that precipitation is a complex phenomenon with many spatial features. Second, the model, having been specifically trained for this task, is able to capture the spatio-temporal dependencies present in the dataset thanks to its non-linear, convolutional structure.

Other researchers, such as Hall et al. (2022) and Moreira et al. (2022), have also confirmed the performance of ConvLSTM. These results underline the ability of the ConvLSTM model to capture the spatial dependencies between input variables (precipitation, temperature, and wind) linked to flow prediction. Given that these variables are highly complex, with both spatial and temporal variations, taking into account their spatial dimensions has a significant influence on discharge at a watershed outlet. This is why it is so important to have a model capable of capturing these dependencies to improve prediction accuracy. The results detailed in this section demonstrate the considerable impact of AI and satellite data on improving flood forecasting. In-depth satellite data analysis combined with advanced AI techniques has revealed good accuracy, paving the way for more effective risk management strategies. While recognizing the continuing challenges, it is clear that the synergy between these advanced technologies represents considerable potential to revolutionize flood prediction and management in the future.

This study examined the application of AI and remote sensing in flood forecasting, focusing specifically on the Ouémé basin at the Bétérou outlet. Our results clearly showed that the use of advanced ML techniques, coupled with high-precision satellite data, can significantly improve the reliability of flood forecasts. The ConvLSTM model, which merges the spatial analysis of CNN with the memory capacity of LSTMs, has proved exceptionally effective in flood forecasting with the Nash and R² about 0.96. This breakthrough represents a major advance, providing a more nuanced understanding of complex hydrological dynamics and equipping decision-makers with innovative tools for proactive flood risk management, particularly in communities most exposed to hydrological risks. Although this research has made significant progress, there are still challenges to be overcome, such as improving the generalization of models to unexpected scenarios and taking account of uncertainties in predictions. However, the promising results provide strong encouragement for further research in this crucial area, where AI and remote sensing are playing an increasingly prominent role in preventing natural disasters.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.