The research, led by Satti J. R. Kamenan of the National Polytechnic Institute Félix Houphouët-Boigny in Yamoussoukro, is the first attempt at near real-time rainfall prediction in any of Ivory Coast’s river basins. The team built a forecasting framework around Long Short-Term Memory neural networks, a deep learning architecture specifically designed to remember patterns across long sequences of data. LSTM networks, first proposed by Sepp Hochreiter and Jürgen Schmidhuber in 1997, use internal gates—forget, input and output gates—that regulate what information is stored, discarded or passed forward through the network’s memory cells. This structure allows them to capture the long-term temporal dependencies and nonlinear behavior that characterize daily rainfall, something conventional statistical models have long struggled with.

Feeding the neural network required an ambitious data pipeline. The researchers combined ground-based rainfall records from SODEXAM, the Ivorian meteorological agency, covering ten stations in the Marahoué watershed from 1961 to 2022, with two NASA satellite products: GPM IMERG, a high-resolution precipitation dataset with a spatial resolution of 0.1 degrees and roughly 30-minute temporal resolution, and MERRA-2, a reanalysis that blends satellite observations, ground measurements and numerical weather models into consistent atmospheric fields including temperature, wind, pressure and humidity. All of this was processed on Google Earth Engine, the cloud-based platform that has become a workhorse for large-scale environmental data analysis. Before the satellite data could be trusted, the team validated it against rain gauge observations from 2000 to 2017, finding correlation coefficients of 0.80 at Bouaflé and 0.83 at Zuénoula—meaning the satellite product captured more than 80 percent of the observed rainfall variability.

Quality control was a critical and methodical step. The historical gauge records contained gaps of 7.29 to 8.04 percent, and raw observations can harbor errors that would corrupt any machine learning model. The team applied a two-stage screening protocol: first, the interquartile range method flagged values that deviated unusually far from each station’s typical distribution; second, a spatial consistency check using the network-wide median and median absolute deviation determined whether a flagged value was physically plausible or simply wrong. Crucially, extreme rainfall events that were spatially coherent across the network were retained even when they exceeded statistical thresholds, preventing the accidental deletion of genuine meteorological extremes. Only 689 of 29,510 observations—2.34 percent—were ultimately identified as outliers, and erroneous values were corrected using the validated GPM IMERG estimates as an auxiliary source.

Selecting the right predictors was equally deliberate. From twenty-one candidate atmospheric variables, the researchers used a Pearson correlation matrix to eliminate multicollinearity, keeping only variables with pairwise correlations below 0.60. They then ran a Random Forest model to rank the remaining candidates by importance, retaining twelve predictors—including seven-day antecedent rainfall, vertical atmospheric motion at 500 hectopascals, specific humidity at multiple pressure levels, and wind components—that each contributed meaningfully to predictive accuracy. The final dataset spanned 15,094 daily time steps from January 1984 to April 2024, with data through January 2020 used for training and the final four years reserved entirely for independent validation, a rigorous test of whether the models had truly learned generalizable patterns rather than memorizing the past.

Three separate LSTM models were trained, one for each forecast horizon: one day ahead, three days ahead and seven days ahead. The architecture settled on three layers of 100 neurons each, a learning rate of 0.001 with automatic reduction, a batch size of 32, and an early stopping mechanism that halted training when validation loss stopped improving. The team then pitted the LSTM against three popular machine learning benchmarks—Random Forest, Extra Trees and XGBoost—using standard metrics including the coefficient of determination, Nash-Sutcliffe Efficiency, normalized root mean square error and mean absolute error. The results were striking. At the one-day horizon, the LSTM achieved R-squared and NSE values exceeding 95 percent at both stations, with a normalized RMSE below 10 percent and a mean absolute error under 1 millimeter, placing it firmly in the ‘excellent’ category by conventional hydrological standards.

The gap between the deep learning model and its competitors widened as the forecast horizon stretched. At three days ahead, the LSTM still maintained R-squared values between 88 and 90 percent with a normalized RMSE below 10 percent, while the tree-based models saw their accuracy slip, with R-squared often dropping below 65 percent. By seven days ahead, the LSTM was the only model still producing statistically reliable forecasts, with R-squared values between 70 and 81 percent and errors remaining controlled, whereas Random Forest and Extra Trees deteriorated dramatically—mean absolute errors exceeding 7 millimeters and NSE values sometimes falling below 40 percent—and XGBoost performed worst of all. The explanation lies in fundamental architecture: tree-based ensembles treat observations largely as static feature sets, while the LSTM’s recurrent memory cells explicitly model the sequential structure and long-range dependencies inherent in daily precipitation series.

Honest scrutiny of the model’s errors revealed both strengths and limits. Residual analysis showed that predictions were largely unbiased for light to moderate rainfall, which constitutes the majority of events in the Marahoué basin, with mean residuals near zero at both stations. But the Kolmogorov-Smirnov test confirmed the residuals were not normally distributed, and a class-by-class examination exposed an intensity-dependent bias: some low-intensity events were slightly underestimated while moderate and heavy rainfall tended to be overestimated, with error dispersion growing sharply for extreme events. This is a well-documented weakness of deep learning applied to precipitation—intense rainfall is rare, irregular and heavily skewed, so models see too few examples to learn its behavior precisely, and recent studies have shown LSTM networks tend to smooth high-intensity peaks rather than reproduce them faithfully.

For the people of Bouaflé and Zuénoula, however, the practical implications are significant. Early warning systems depend less on perfectly reproducing once-a-decade extremes than on reliably tracking the ordinary rainfall regime and flagging major events in time for communities to act—and the LSTM delivers exactly that at lead times of one to three days, with usable guidance even at seven. The authors point to clear paths forward: enriching the training data with additional atmospheric predictors such as convective available potential energy, total column water vapor and sea surface temperatures over the Gulf of Guinea and ENSO regions, applying post-prediction bias correction, and exploring hybrid models that couple neural networks with physical approaches for extreme events. As climate change intensifies rainfall variability across West Africa, this study demonstrates that a developing region can build world-class forecasting capability from open satellite data, cloud computing and a well-designed neural network—no expensive radar infrastructure required.

Subject of Research: Deep learning-based daily rainfall forecasting using satellite and reanalysis data for flood early warning in the Marahoué River basin, Ivory Coast

Article Title: Intelligent daily rainfall prediction for early warning using deep learning and satellite data: application to Bouaflé and Zuénoula stations, Ivory coast

Article References: Kamenan, S. J. R., Youan, T. M., Adja, M. G., Soro, S. I., & Kouassi, A. M. (2026). Intelligent daily rainfall prediction for early warning using deep learning and satellite data: application to Bouaflé and Zuénoula stations, Ivory coast. Advances in Statistical Climatology, Meteorology and Oceanography, 12(1), 173-193. https://doi.org/10.5194/ascmo-12-173-2026

Image Credits: AI Generated

DOI: 10.5194/ascmo-12-173-2026

Keywords: deep learning, LSTM neural networks, rainfall prediction, flood early warning, GPM IMERG, satellite data, Ivory Coast, Marahoué River basin, machine learning, hydrology, MERRA-2, extreme rainfall

The stakes are far from academic. The Western Region, wedged between Côte d’Ivoire and the Gulf of Guinea, receives between 1,500 and 2,000 millimetres of rain each year, sustaining cocoa farms, timber plantations and oil palm groves that anchor the local economy. Its coastal capital, Sekondi-Takoradi, has repeatedly seen heavy rainfall overwhelm drainage systems, and flooding remains a chronic threat. Better forecasts translate directly into better decisions about when to plant, when to harvest, how much water to store, and when to warn communities that a flood may be coming.

The researchers drew on validated monthly records from the Ghana Meteorological Agency spanning January 2017 to December 2023, with individual monthly totals ranging from zero to 419 millimetres. Before any modelling began, they confirmed that the series was statistically stationary using both the augmented Dickey-Fuller test, which returned a statistic of -4.34 with a p-value of 0.01, and the KPSS test, which failed to reject stationarity at the 0.1 level. That dual confirmation matters because models like SARIMA assume a stable mean and variance; if the underlying data drift, the forecasts inherit the drift as error.

Extreme values posed a subtler problem. A pronounced rainfall spike around 2020, likely tied to extreme weather, threatened to distort parameter estimates for every model. The team tested two remedies: winsorization, which caps extreme values at the 5th and 95th percentiles while keeping them in the dataset, and the interquartile range method, which effectively trims outliers away. Winsorization proved the better tool. It preserved the central tendency of the data, holding the mean at 94.24 millimetres against the original 97.43, while taming the standard deviation from 95.93 to 86.65. The IQR method pushed the mean down to 83.67 and, crucially, discarded information about genuine extreme events. For a forecasting problem where the extremes are precisely what you want to anticipate, that distinction is decisive.

Prophet’s architecture explains much of its advantage. The model decomposes a time series into a trend component, a seasonal component modelled as a Fourier series, a holiday or event component, and a noise term. Its trend can follow either a piecewise linear path with automatically detected change points or a logistic growth curve with a carrying capacity, allowing it to bend when the data bend rather than forcing the data into a fixed seasonal structure. SARIMA, by contrast, relies on autoregressive and moving average terms with rigid seasonal orders, which the authors note struggles with the nonlinear, abrupt shifts characteristic of tropical rainfall. ETS smoothing methods and TBATS, which handles multiple seasonal cycles through trigonometric terms, fared better in some respects but still lagged behind.

The performance gap was substantial. Prophet achieved the lowest Akaike information criterion at 880.09 and Bayesian information criterion at 884.95, alongside a mean squared error of 1,981.79 and a mean absolute error of 35.15. SARIMA, the traditional workhorse, posted an AIC of 947.18, an MSE of 3,590.58 and an MAE of 46.15. The hybrid ARIMA-plus-ETS model landed in between, with an AIC of 1,005.93, while TBATS and ETS trailed further. On explanatory power, Prophet reached an R-squared of 0.92 with a mean absolute percentage error of 4.8 percent, compared with 0.85 and 7.3 percent for SARIMA, 0.79 and 8.6 percent for ETS, and just 0.76 and 9.2 percent for TBATS. When winsorization was applied, Prophet’s R-squared climbed from 0.89 to 0.92 and its MAPE fell from 5.6 to 4.8 percent, whereas IQR-based removal degraded performance by discarding data.

Residual diagnostics reinforced the case. A Box-Ljung test on the model’s residuals returned a p-value of 0.1055, indicating no significant autocorrelation and suggesting the model had captured the systematic structure of the series. Standardized residual plots showed most errors clustering near zero, though spikes around 2019 and 2021 hinted at anomalies, possibly extreme weather events, that no model fully absorbed. The fitted model’s parameters told a story of a rainfall regime in slow flux: a small positive growth rate of roughly 0.00036 per month, negligible seasonal coefficients, and change-point values so small that the authors concluded the overall trend has remained remarkably stable despite year-to-year drama.

Looking forward, the model’s 2024 forecast projected the familiar bimodal pattern with peaks exceeding 200 millimetres in May and June, including a May prediction of 224.47 millimetres within a confidence interval stretching from 173.18 to 279.09 millimetres. On the fitting sample, Prophet posted an R-squared of 0.9999895, a Theil’s U statistic of 0.0013 and a MAPE of just 1.116 percent, figures the authors interpret as evidence that the model captures nearly all variance without overfitting, since the adjusted R-squared differed only marginally. The wide confidence intervals during peak months, however, are a reminder that uncertainty balloons exactly where the consequences are largest.

The authors are candid about limitations. Prophet assumes a degree of climate stationarity, the expectation that historical patterns will persist, which accelerating climate change may erode. Monthly data, moreover, can hide the short-duration cloudbursts that matter most for flash flood warning. They argue that future systems should fuse Prophet with machine learning anomaly detection, such as isolation forests or neural classifiers, and with high-resolution geospatial data to move from regional averages to village-scale advisories. Deploying such tools nationally would demand investment in training, real-time data pipelines and infrastructure, costs the authors frame as small against the losses from missed floods and mistimed harvests.

What emerges is a quietly radical suggestion for operational meteorology in data-scarce regions: the most flexible tool won, and it won by respecting the messiness of the data rather than disciplining it. Earlier work in the same region using an ARMA model explained only 53 percent of rainfall variation; even a recent SARIMA effort with careful outlier adjustment reached an R-squared of 99.02 percent on its fitting sample but required extensive manual preprocessing. Prophet’s automated handling of trend shifts, seasonality and outliers offers a scalable path for Ghana’s National Climate Change Policy ambitions, from flood early warning systems in Takoradi to irrigation scheduling across the cocoa belt. As the authors put it in their conclusion, models that generalize well and account for nonlinearity and seasonality are the ones that will help climate-sensitive societies stay ahead of the rain.

Subject of Research: Comparative time series modelling of monthly rainfall forecasting in Ghana's Western Region

Article Title: Advanced time series modelling to address seasonal variability and extreme rainfall in Ghana

Article References: Bosson-Amedenu, S., Baah, E. M., Ayiah-Mensah, F., & Addor, J. A. (2025). Advanced time series modelling to address seasonal variability and extreme rainfall in Ghana. BMC Environmental Science, 2(1), Article 8. https://doi.org/10.1186/s44329-025-00024-8

Image Credits: AI Generated

DOI: 10.1186/s44329-025-00024-8

Keywords: rainfall forecasting, Facebook Prophet, SARIMA, time series analysis, Ghana, Western Region, outlier handling, winsorization, climate resilience, seasonality, flood early warning, BMC Environmental Science

The problem the researchers set out to solve is one that hydrologists and machine learning engineers have been circling for years. Deep learning models have repeatedly proven that they can forecast river levels and streamflow with impressive accuracy, often outperforming classical hydrological simulation in data-rich settings. But accuracy alone is not enough for early warning. Many state-of-the-art networks carry enormous parameter counts and require long inference times, meaning the interval between feeding in the latest gauge readings and receiving a prediction can stretch unacceptably long. At the same time, standard recurrent architectures tend to smooth over exactly the signals that matter most during a disaster: the abrupt, nonlinear jumps in water level that mark the onset of an extreme flood event. A model that averages away a flood peak is, for warning purposes, worse than useless.

BOA-LSTM attacks both weaknesses simultaneously through three coordinated design choices. First, the authors deliberately reduce the number of LSTM layers and the number of hidden units within each layer. Long short-term memory networks, introduced by Sepp Hochreiter and Jürgen Schmidhuber in 1997, use gated cells to preserve information over long sequences, but stacking many such layers multiplies parameters and computation. By trimming the architecture to the minimum needed for hydrological time series, the team cut computational complexity and slashed inference time, the critical metric for real-time deployment on modest hardware at remote monitoring stations.

Shrinking a network, however, risks throwing away its sensitivity to subtle patterns, so the second design choice compensates by adding a scaled dot-product attention mechanism. Attention layers, which rose to fame in natural language processing, allow a model to dynamically weight the most informative parts of an input sequence rather than treating all time steps equally. In the flood forecasting context, this means the network can amplify sudden rises in discharge or water level in the recent past, the abrupt hydrological changes that precede extreme events, instead of letting them dissolve into the longer record of routine variability. The attention weights effectively teach the model where to look when a river begins to misbehave.

The third pillar is Bayesian optimization, a sample-efficient strategy for tuning hyperparameters such as learning rates, hidden unit counts, and window sizes. Traditional hyperparameter search relies on grid or random searches that evaluate huge numbers of configurations, an expensive process for any deep learning pipeline. Bayesian optimization, famously formalized for machine learning by Snoek, Larochelle, and Adams in 2012, builds a probabilistic surrogate of the performance landscape and intelligently selects which configurations to test next. In BOA-LSTM this automates the search for key hyperparameters, improving configuration efficiency and removing a significant degree of manual trial and error from the modeling workflow, which matters for agencies that cannot employ teams of specialists to hand-tune every new river basin.

The empirical results are the heart of the study’s claim. Using the USGS 02337000 benchmark dataset, a river gauging record from the United States Geological Survey network, the team showed that BOA-LSTM achieves a mean squared error of 0.000295 with an inference time of 124 seconds. Benchmarked against FAIRDNN, a comparable deep learning approach, the lightweight model delivered lower overall prediction errors, a notably better representation of flood peaks, and substantially higher inference efficiency. The numbers are striking: mean squared error fell by 4.8 percent while inference time dropped by 81.8 percent. In practical terms, the model is simultaneously more accurate and roughly five times faster than its competitor, a rare combination in machine learning where speed and precision usually trade off against each other.

The improved flood-peak representation deserves particular emphasis, because peaks are where forecasting models most often fail and where failure is most costly. Extreme events are by definition rare, so they are underrepresented in training data, and loss functions dominated by routine conditions can teach a network to systematically underestimate the highest stages. By reducing overall error while sharpening the model’s attention to abrupt changes, the architecture appears to preserve precisely the tail behavior that converts a generic water level forecast into an actionable warning. For emergency managers, an underpredicted crest of even a few centimeters can mean the difference between a precautionary evacuation and a rescue operation.

To demonstrate that the approach is not tied to a single American gauging station, the researchers conducted a case study using 2024 water level observations from Cambridgeshire in the United Kingdom. This second dataset exercised the full operational pipeline on data from a different hydrological setting, including the warning classification process in which predicted levels are translated into graduated alert categories. The case study, the authors report, demonstrates the model’s applicability to other stations and provides a reference workflow for how raw predictions become graded warnings, the step where machine learning output meets the public communication machinery of civil protection agencies.

The broader context makes the work timely. Flood risk is intensifying in many regions as climate change amplifies rainfall extremes and as development pushes populations into floodplains, while the scientific literature has seen a rapid proliferation of attention-enhanced and Bayesian-assisted forecasting models, from dual-stage attention LSTMs for multi-step flood prediction to hybrid architectures for data-scarce mountain catchments and Bayesian deep learning frameworks for uncertainty estimation. What distinguishes BOA-LSTM within this crowded field is its explicit commitment to lightweight deployment, treating inference speed and parameter economy as first-class objectives rather than afterthoughts. That orientation reflects a growing recognition that a forecasting model earns its keep only when it runs reliably on the infrastructure actually available in the field.

Caveats remain, as they do with any data-driven forecasting system. The model’s performance was established on benchmark and case study datasets rather than through live operational trials during an actual flood emergency, and the published data availability statement indicates that no new datasets were generated or analyzed during the study beyond those used in the experiments. Deep learning forecasters also inherit the limitations of their training records: under nonstationary climate conditions, tomorrow’s extremes may look statistically unlike yesterday’s, a challenge the wider hydrology community continues to grapple with. Even so, the study offers a concrete demonstration that flood warning need not choose between intelligence and immediacy. With an 81.8 percent reduction in inference time and improved error performance, BOA-LSTM suggests that the next generation of flood early warning systems may be both smarter and lighter, bringing life-saving minutes to communities that live downstream of the world’s increasingly volatile rivers.

Subject of Research: A lightweight LSTM model using Bayesian optimization and attention mechanisms for real-time extreme flood early warning

Article Title: BOA-LSTM: a lightweight LSTM model integrating bayesian optimization and attention mechanism for real-time extreme flood early warning

Article References: Zhang, Y., Zhou, M., Chen, M., & Jia, H. (2026). BOA-LSTM: a lightweight LSTM model integrating bayesian optimization and attention mechanism for real-time extreme flood early warning. International Journal of Machine Learning and Cybernetics, 17(10), Article 477. https://doi.org/10.1007/s13042-026-03313-z

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03313-z

Keywords: flood early warning, BOA-LSTM, LSTM, Bayesian optimization, attention mechanism, machine learning, hydrology, time series forecasting, deep learning, USGS, flood prediction, real-time systems