Accurate river discharge forecasting is critical for flood mitigation and effective water resource management, particularly in data-scarce regions. Capturing both spatial rainfall variability and temporal discharge dynamics remains a challenge.
Methods
An innovative hybrid forecasting model was developed to predict hourly river discharges up to 24 h in advance. The proposed model combined the spatial analysis capabilities of convolutional neural networks with the temporal sequence modeling strengths of long short-term memory networks. It was designed to extract spatio-temporal patterns from satellite precipitation data while incorporating historical discharge trends to improve forecasting accuracy. The model was developed and internally validated on a 3 day lag dataset before being tested on an external test set with a relatively shorter 30min lag to simulate real-time conditions.
Results
The hybrid CNN-LSTM model demonstrated strong predictive performance across multiple evaluation metrics, including the correlation coefficient, Nash-Sutcliffe efficiency, root mean square error, and mean absolute error. Notably, our proposed model maintained a performance level well within acceptable error limits, even using the 24 h prediction horizon. The results confirmed the model's ability to generalize across datasets with differing temporal resolutions.
Discussion
The proposed framework provides a scalable and robust solution for improving short-term discharge predictions, particularly in upstream and unmeasured catchments where ground observations are limited. This approach has strong potential to support real-time water resource management and flood preparedness in data-scarce regions.
convolutional neural networksflood forecastinghourly discharge datalong short-term memoryprecipitation satellite dataThe author(s) declared that financial support was received for this work and/or its publication. This research was supported by the Kasetsart University Research and Development Institute (KURDI), Bangkok, Thailand under the “FF(KU)36.69” project (to WT).section-at-acceptanceWater and Artificial IntelligenceIntroduction
The impacts of climate change have intensified extreme weather patterns globally, resulting in heavier rainfall, stronger cyclones, more intense heatwaves, and prolonged droughts. In 2023, Asia was the most disaster-affected region in the world due to extreme weather conditions. Cyclones and floods were the leading causes of fatalities and economic losses, with 79 meteorological disasters reported across the region. Over 80% of these events were attributed to floods and storms, resulting in more than 2,000 deaths and affecting over 9 million people (WMO, 2024). Early warning systems play a critical role in mitigating the devastating impacts of such hazards. These systems provide timely, relevant, interoperable, accessible, and accurate information to support preparedness and response efforts. Notably, disaster-related damage can be reduced by up to 30% if warnings are issued at least 24 h in advance (WMO, 2022) Flood forecasting is essential for early warning systems to develop appropriate measures to control flood risks, mitigate hazards, evacuate vulnerable populations from flood-prone areas, set insurance premiums, and optimize environmental and water resource management.
Over the past decade, deep learning (DL) techniques have emerged as a powerful tool for tackling complex hydrological challenges, particularly in flood forecasting. These techniques leverage their ability to handle sequential input data and capture the non-linear dependencies inherent in hydrological processes (Kumar et al., 2023; Shen et al., 2021; Sit et al., 2020). This development reflects a general trend toward the integration of data-driven methods into traditional hydrological domains. Recent advancements have introduced a range of DL architectures specifically tailored for flood forecasting. Recurrent neural networks (RNNs) and their derivatives, including gated recurrent units (GRUs) and long short-term memory networks (LSTMs), have shown promise in modeling temporal dependencies in flood-related data (Waqas and Humphries, 2024). In a notable study by Kratzert et al. (2018), LSTMs have been used to simulate rainfall-runoff processes and showed superior performance compared to classical hydrological models in several catchments. Hydrological parameters have been incorporated into the learning process to improve the accuracy of DL models. These integrations enable a more comprehensive understanding of complex hydrological systems (Chang et al., 2023; Senent-Aparicio et al., 2019). Tongal and Booij (2018) proposed a framework for streamflow simulation and prediction, demonstrating improved predictive performance by decomposing streamflow into individual components. Thaisiam et al. (2022) introduced the concept of concentration time (TC) to enhance temporal correlations using backward lags. Integrating TC-related features with RNNs and their variants essentially demonstrated superior performance in multi-step hourly flood prediction. Several studies have investigated seasonal overflow of water, a phenomenon frequently observed during wet seasons and flood events (Lee et al., 2020; Thaisiam et al., 2024; Zanchetta et al., 2022). Zhang and Yan (2023) proposed LSTM-based network components to address long-term and short-term dependencies separately, while Chenmin et al. (2024) investigated feature selection techniques for a hybrid model using convolutional neural networks and bidirectional LSTMs with an attention mechanism for flood prediction. Overall, DL techniques have demonstrated exceptional efficiency in flood forecasting, particularly in study areas with long-term and continuous monitoring data.
Convolutional Neural Networks (CNNs) are valuable tools for analyzing satellite images because they are particularly adept at handling spatial data. For example, their ability to extract and analyze spatial features enables a deep understanding of patterns in satellite data, such as precipitation intensity and distribution. By learning localized and large-scale spatial dependencies, CNNs enable a more detailed representation of hydrometeorological processes compared to traditional approaches. The recent trend in flood forecasting research involves combining CNNs with other DL techniques to address varied challenges of predicting floods. These hybrid approaches aim to leverage the strengths of different methodologies, such as spatial data modeling via CNNs and temporal sequence prediction via recurrent models (Wu et al., 2020). For example, Feng et al. (2022) utilized hybrid models combining spatio-temporal graph convolutional networks and LSTMs to forecast flood levels, highlighting the potential of integrating spatial and temporal modeling. These approaches effectively bridged the gap between static and dynamic flood prediction factors.
In recent years, CNN–LSTM architectures have been applied to rainfall–runoff modeling by integrating meteorological forcing with hydrological observations from upstream catchments. For example, Li X. et al. (2022) introduced a CNN-LSTM based rainfall-runoff model that utilized observed precipitation represented as two-dimensional spatial inputs to predict downstream flow, while Li P. et al. (2022) incorporated radar maps together with upstream runoff observations. Later, a hybrid model based on CNN-LSTM and GRU with the improved sparrow search algorithm was proposed for runoff prediction using meteorological, hydrological, and runoff observations (Yao et al., 2023). Wang et al. (2023) proposed a CNN-LSTM model in which CNNs extracted physical and meteorological features from weather research and forecasting model outputs, and the resulting time-series features were used by LSTMs to simulate streamflow dynamics. Additionally, Malik et al. (2024) utilized a time-distributed CNN-LSTM model to improve the accuracy of flood predictions, particularly in forecasting flood occurrence times and peak intensities using station-based precipitation and streamflow data. Hu et al. (2024) used a CNN-LSTM model that integrated multiple grid-based meteorological data with gauged precipitation and runoff sequence inputs. Collectively, these studies demonstrate the effectiveness of CNN–LSTM hybrid models for streamflow simulation and discharge prediction when rich observational data are available. However, despite their demonstrated performance, most existing CNN-LSTM based runoff forecasting studies rely on various ground-based observations from upstream areas, such as rain gauge data, upstream discharge stations, or hydrometeorological information, which limits their applicability in regions with sparse monitoring networks. In contrast, our proposed framework integrates satellite-based precipitation imagery with discharge observations from only the target prediction station, thereby reducing the dependence on upstream ground-based data. While satellite-based precipitation products provide extensive spatial coverage and consistent temporal resolution, their integration into deep learning–based runoff forecasting frameworks remains relatively underexplored. Few studies have explicitly exploited the spatial structure of satellite-derived rainfall fields as primary model inputs, and uncertainties inherent in satellite observations are not fully addressed. Consequently, developing reliable streamflow prediction models for flood events in data-scarce catchments—particularly with limited upstream ground-based information— remains a challenge, and the practical potential of CNN–LSTM architectures for discharge forecasting under such conditions has not been fully demonstrated.
In response, the main objective of this study was to develop a robust model for predicting hourly river discharge up to 24 h in advance using only satellite-based precipitation and target discharge records. The proposed model integrated the spatial analysis capabilities of CNNs with the temporal sequence modeling strengths of LSTMs. This hybrid architecture was designed to extract meaningful spatio-temporal patterns from satellite precipitation data while incorporating historical discharge trends to enhance forecast accuracy. The proposed model utilized spatial precipitation data from the Global Satellite Mapping of Precipitation (GSMaP), which are produced and distributed by the Earth Observation Research Center (EORC) of the Japan Aerospace Exploration Agency (JAXA), as comprehensively reviewed by Kubota et al. (2020). The GSMaP dataset offers high-resolution, real-time precipitation measurements with broader spatial coverage than conventional ground-based sensors. The satellite data were complemented by historical discharge data from a downstream observation station, allowing the model to account for localized hydrological dynamics. The hybrid model demonstrated the ability to generalize across different datasets by combining CNNs for spatial feature extraction with LSTMs for temporal sequence analysis. Internal validation using GSMaP-MVK data assessed the reliability of the model within the training domain, while external tests with GSMaP-NOW data evaluated its performance under real-time conditions. The correlation coefficient (R), Nash-Sutcliffe efficiency (NSE), root mean square error (RMSE) and mean absolute error (MAE) were used to rigorously evaluate the predictive capabilities of the model.
Our main contribution highlighted the potential of hybrid deep learning approaches to enhance hydrological forecasting, particularly in data-scarce regions. We established a robust foundation for the development of advanced deep learning architectures by integrating satellite-based precipitation measurements with station-based discharge data. This integration not only enhanced the accuracy of the discharge predictions but also offered a scalable solution for improving water resource management, especially in catchments with limited upstream ground-based observations. Our proposed framework was rigorously evaluated across two distinct study areas to validate its robustness and generalizability. Furthermore, we implemented a training and validation strategy designed to account for differences in satellite precipitation latency between offline model development and near-real-time forecasting. The model was trained and internally validated on a dataset characterized by a relatively longer time lag, then, it was applied to a different dataset with a shorter time lag. This approach was specifically designed to optimize the model's readiness for real-time deployment, demonstrating its applicability in dynamic hydrological environments.
Materials and methodsStudy area
Figure 1 depicts the two study areas in Thailand: the Upper Yom River Basin and the Upper Chi River Basin. The Upper Yom River Basin is located in northern Thailand and is oriented in a north-south direction. The main river (Yom) originates from Doi Khun Yuam in the Phi Pan Nam mountain range, situated in the Pong and Chiang Muan districts of Phayao province, Thailand. The Yom river flows through a steep valley with a riverbed slope of approximately 1:120 and an elevation range of 180 to 360 m above mean sea level. Narrow plains are present along certain sections of the riverbank before the river flows into Phrae province, Thailand. Flooding in this basin occurs annually during the rainy season due to the absence of large water storage facilities and the steep terrain. These factors contribute to rapid runoff, which causes river overbank flows in downstream areas.
Study areas equipped with stream gauges for (a) The upper Yom River Basin, (b) The upper Chi River Basin.
Topographic map showing two river basins in Thailand with their district centers and stream gauges marked, province boundaries outlined, and elevation represented by color shading; inset map highlights their location within Thailand.
The topography of the Upper Chi River Basin consists of high mountain ranges. The Phu Phan Mountain Range lies to the east and north, while the Dong Phaya Yen Mountain Range, located to the west, serves as the source of the Chi River. The Chi River has a riverbed gradient of 1:630 in its upstream area. Additionally, the Chi River is fed by tributaries, including the Phrom River, Sapung River, Chiang Tha River, and Khan Chu River, respectively. The combination of the basin's topography and its hydrological system results in heavy rainfall in the upstream areas, causing simultaneous water flow from the main river and its tributaries toward Mueang district, Chaiyaphum province, Thailand. This often leads to flash floods in downstream areas. Key characteristics of study areas are summarized in Table 1.
Characteristics of study areas.
Characteristics
Upper Yom River basin
Upper Chi River basin
Observation station
Y.1C
E.21
Annual runoff (million cubic meters)
1,982
1,420
Data period
1998–2024 (27 Years)
2005–2024 (20 Years)
Watershed (square kilometers)
7,660
9,755
River length to station (kilometers)
180
273
Average river slope (kilometers/kilometers)
1:120
1:630
Time of concentration, Tc (hour)
50
48
Methodology
This study aimed to develop a robust model for predicting hourly discharge up to 24 h in advance, applicable to poorly gauged areas where traditional upstream observed data are unavailable. Satellite-based precipitation data served as the primary input for the model, complemented by historical station-based data of the target variable. A hybrid deep learning architecture was constructed to effectively capture the spatiotemporal relationships within the data, combining the strengths of CNNs and LSTMs. Preprocessed satellite data, along with historical discharge data, were used to train and fine-tune the model. A rigorous internal validation process, leveraging data from the same source, was conducted to evaluate the model's performance within the training domain. Subsequently, the model's generalizability was assessed by validating it on an external test set sourced from a different satellite source. A flow chart in Figure 2 illustrates the entire process, consisting of data acquisition, preprocessing, model training, and evaluation.
Process flowchart of (a) data pre-processing, (b) model development, (c) model evaluation, and (d) external testing.
Flowchart depicting a machine learning workflow for hydrological modeling, including data pre-processing of precipitation images and water levels, model development using CNN and LSTM layers, model fitting, evaluation with performance indices, external testing, and conclusion.Data pre-processing
Our primary objective was to develop a multi-step discharge prediction model for the downstream water level station. To achieve this, we initially preprocessed hourly water level data to compute more informative discharge values based on an annual rating curve. A traditional outlier detection method was applied to eliminate extreme anomalies from the dataset. Specifically, summary statistics including an interquartile range and time-series plots were examined to identify anomalous values. Extreme values that were consistent with known high-flow or intense rainfall events were retained, as such extremes represented meaningful hydrological behavior and were critical for evaluating model performance under flood conditions. The resulting discharge data at the downstream station served a dual role in the model, acting as both the prediction target and input features representing historical flow conditions. Due to the limited availability of upstream observed hydrological data, we incorporated satellite-based data to monitor hydrological patterns. Specifically, we utilized data from GSMaP, a network of geostationary weather satellites managed by JAXA. It provides real-time weather information, including precipitation, cloud cover, and atmospheric conditions, across the Asia-Pacific region. These data are generated from dual-frequency precipitation radar measurements collected by multiple satellites.in
For our model, we relied on hourly GSMaP precipitation data with a spatial resolution of 0.1 degrees (approximately 11.1 km), striking a balance between spatial coverage and fine-grained data granularity. The primary dataset used for model training was GSMaP-MVK, which incorporates a 3 day time lag. This dataset was selected for its reliability and the extensive availability of historical data, making it well-suited for developing a robust and accurate predictive model. To assess the model's real-time performance, we integrated GSMaP-NOW, a dataset offering precipitation data with a considerably reduced time lag of 30 min. This near-real-time data source enabled us to validate the model's capability in operational settings, facilitating its ability to deliver timely and actionable predictions for discharge scenarios.
After collecting and preprocessing the data, the GSMaP-MVK dataset was divided into three subsets for training, validation, and testing. The training set was used to fit the model, while the validation set facilitated fine-tuning of the model and mitigating the risk of overfitting. Then, the final model was evaluated on a holdout test set. The study focused on two key river basins: the Upper Yom River and the Upper Chi River. Validation utilized data from a 3 year period (2016–2018), while the testing phase utilized data spanning 2019 to 2021. The training set covered 18 years of data (1998–2015) for the Upper Yom River basin and 11 years (2005–2015) for the Upper Chi River basin. The model's performance was further assessed based on recent data, using an external test set comprising GSMaP-MVK and GSMaP-NOW data from 2022 to 2024. An overview of the discharge data for both study areas, highlighting the temporal patterns and variability in water discharge observed during the study period, is provided in Figure 3.
Data components for training, validation, testing and external testing for (a) Y.1C station in Upper Yom River Basin and (b) E.21 station in Upper Chi River Basin based on GSMaP-MVK and GSMaP-NOW data.
Two line graphs compare river discharge in cubic meters per second for stations Y.1C and E.21. Both graphs are split into shaded periods indicating training, validation, testing, and extended testing phases. Above the charts, a timeline marks the availability of GSMaP-MVK since 1998 and GSMaP-NOW since December 2021. The first graph (Y.1C) spans 18 years of training data, and the second graph (E.21) covers 11 years. Color-coded sections and labels highlight sequence and purpose of data periods.
To preprocess the satellite image data, we focused on extracting relevant grid data corresponding to the study areas, specifically targeting regions within the river basin boundaries. Regions outside these boundaries were excluded to ensure that the model processed only information pertinent to the river basins of interest. The zero-padding technique was applied to maintain a consistent input size of 16 by 16 pixels. Furthermore, instead of reshaping the entire image, which could distort spatial relationships within the data, zero-padding was used to fill empty grid cells with zero values. This approach preserved the integrity of spatial features while standardizing the image dimensions. According to Hashemi (2019), the zero-padding technique had no negative impact on the model accuracy but essentially reduced the training time. The padding ensures uniform input dimensions without compromising the representation of hydrological features. A fixed input size of 16 by 16 pixels was applied across two study areas to ensure compatibility with the CNN architecture. Figure 4 illustrates an example of the satellite image processing workflow, highlighting both the boundary reshaping and the zero-padding technique applied to the Upper Chi River Basin. As shown in Figure 4, white grid cells with actual rainfall values were retained to preserve the spatial structure of satellite precipitation fields. These zero values represent actual dry conditions reported by the satellite product rather than missing or zero-padding grids.
Imagery processing using boundary reshaping and zero padding for E.21 data.
Three-panel diagram illustrating rainfall data processing over a geographical area labeled E.21. The left panel shows raw rainfall intensity with a color scale, the center panel outlines area cropping with a red border, and the right panel displays the result fitted into a 16-by-16 pixel grid with gray zero padding.Model development
Previous studies have applied CNN–LSTM models for flood forecasting by learning from image-based representations of meteorological and hydrological conditions in upstream areas. The input images are typically generated from multiple data sources, including ground-based observations, model-derived products, radar rainfall imagery, and various supplementary data analyses, in order to construct images that effectively capture the factors influencing discharge at the forecasting station. Consequently, the preparation of such input images constitutes a labor-intensive and relatively complex process, which may constrain the practical applicability of these approaches for real-time forecasting. To overcome limitations associated with sparse ground-based observations in upstream areas and the complexity of input data preprocessing in data-scarce regions, we employed a hybrid deep learning model to capture spatiotemporal relationships within the satellite imagery and historical trends of the target variable. Inspired by Song (2022) and Li X. et al. (2022), a time-distributed CNN was used to extract spatial features from satellite-based precipitation images. The resulting CNN-derived feature vectors were then concatenated with historical discharge inputs and passed to the LSTM component to learn temporal dependencies, using the TC concept as a look-back parameter.
In contrast to previous studies, our proposed framework exhibited key differences. Song (2022) employed a CNN-only architecture for runoff simulation, without incorporating recurrent layers to model temporal dynamics. Li X. et al. (2022) relied on CNN-LSTM architectures for rainfall-runoff modeling based on observation-derived rainfall inputs but did not integrate river flow as an explicit model input. From an architectural perspective, while prior studies primarily utilized custom-designed convolutional stacks tailored to specific datasets, we instead adopted a convolutional layer design based on the Visual Geometry Group-16 (VGG-16) architecture, which has demonstrated strong performance in image-related tasks compared to relatively simpler networks. Our simplified version of the VGG-16 model consisted of two consecutive 2D convolutional layers with 64 filters each. We applied a time-distributed wrapper around the convolutional layers to handle temporal sequences. This allowed the same convolutional operations to be applied to each temporal slice of the input data. In particular, we utilized a TC period as a look-back window (Thaisiam et al., 2022), enabling the model to recognize temporal patterns over longer periods. Max-pooling layers were added after the convolutional layers to reduce data dimensionality while retaining spatial information. A series of convolutional layers and max-pooling layers was additionally stacked, with the final output being flattened into a dense layer comprising 512 units.
The CNN-based architecture was designed to learn spatial relationships from satellite imagery while reducing the high-dimensional data into a 1D vector of size 512. To incorporate the historical discharge information, this 1D vector was concatenated with the downstream discharge data, resulting in a combined feature vector of 513 units, aligned with the same TC-period temporal window. Two LSTM layers, each with 16 units, were used to model the temporal relationships within the historical discharge data. LSTMs were chosen for their effectiveness in capturing long-term dependencies in sequential data. In addition, two fully connected layers were added to transform the LSTM outputs into a suitable representation for forecasting. The final layer consisted of 24 output units, corresponding to the desired forecasting horizons. Figure 5 illustrates the proposed network architecture.
Proposed network architecture.
Diagram of a deep learning architecture combining spatial and temporal inputs. Spatial features enter a simplified VGG-16 convolutional neural network with convolutional and max pooling layers. Temporal features are concatenated and passed to a flatten layer, followed by LSTM layers with sixteen units each, and finally fully connected layers producing a twenty-four dimensional output.
Several standard numerical metrics were utilized to evaluate the performance of our model. The correlation coefficient (R) was used to assess the strength of the linear relationship between the predicted values and the actual target variable. To further quantify the model's error, we calculated both the RMSE and the MAE. RMSE provides a measure of the standard deviation of the residuals, reflecting the average magnitude of the errors, while MAE indicates the average absolute deviation from the true values. Additionally, we computed the NSE coefficient to evaluate the model's effectiveness illogical context. The NSE represents the ratio of the residual variance to the variance of the observed data, offering insight into how well the model performs compared to a baseline prediction. The detailed configuration of the proposed CNN-LSTM framework is summarized in Table 2. The table reports the network architecture, key hyperparameters, and training settings adopted in this study, including input dimensions, convolutional and recurrent layer specifications, optimization strategy, and loss function.
Mean absolute error (MAE), Root mean squared error (RMSE)
Results and discussionTesting dataset results
The multi-step hybrid model integrating CNNs and LSTMs was developed to predict discharge data 24 h in advance. We initially evaluated the performance of the proposed model using the GSMaP-MVK dataset for both the Upper Yom and Upper Chi River Basins. To comprehensively assess the model's accuracy and reliability, we calculated R, NSE, MAE, and RMSE at different forecasting horizons (specifically, 1-step, 6-step, 12-step and 24-step intervals), allowing us to observe how prediction errors progressed over time. Table 3 provides a summary of all these metrics based on the internal test set of GSMaP-MVK for both river basins. The inclusion of multiple forecasting horizons evaluated both short-term prediction accuracy and the model's capability to maintain performance over longer periods.
Performance indices obtained from prediction results of testing dataset.
Performance indices
Upper Yom river basin
Upper Chi river basin
T + 1H
T + 6H
T + 12H
T + 24H
T + 1H
T + 6H
T + 12H
T + 24H
GSMaP-MVK
R
0.9926
0.9799
0.9526
0.8786
0.9970
0.9965
0.9954
0.9921
NSE
0.9765
0.9508
0.8983
0.7633
0.9743
0.9737
0.9726
0.9662
MAE (m3/s)
12.861
15.134
17.996
23.173
9.9051
9.8711
9.9351
10.065
RMSE (m3/s)
17.323
25.045
36.019
54.961
15.540
15.737
16.069
17.843
According to Table 3, the proposed model exhibited an overall desirable performance. Specifically, it maintained a performance level well within acceptable error limits even for the 24-h prediction horizon. Overall, the model's predictive performance generally declined as the prediction horizon increased, particularly when evaluated using R, RMSE, and R. However, slight non-monotonic behavior is observed for MAE in certain cases, reflecting differences in metric sensitivity to extreme discharge errors. The non-monotonic behavior is attributable to its equal weighting of errors, whereas RMSE and NSE are more sensitive to large errors associated with peak flows, and R primarily reflects temporal pattern alignment rather than error magnitude. Across the two study areas, the results for the Upper Chi River Basin consistently outperformed those for the Upper Yom River Basin, achieving R and NSE values exceeding 0.95 for the 24-h forecasting horizon. This disparity could be attributed partially to the differences in the discharge patterns between the two basins. In the Upper Chi River Basin, the topography and river basin system promote a more synchronized flow of water from the main river and its tributaries toward the outlet. This results in relatively more stable discharge patterns with well-defined peaks occurring at specific intervals during the wet season of 2020 and 2021. In contrast, the Upper Yom River Basin, characterized by the feather-like pattern of its watercourses and steep upstream topography, displays highly variable discharge patterns, including abrupt peaks and irregular fluctuations, which make accurate prediction more challenging. These trends are visually represented in Figures 6, 7 for particular flood events, whereas Supplementary Figures 1, 2 in the Supplementary Material show the results for all test periods in the Upper Yom and Upper Chi River Basins, respectively.
Testing dataset results for the Upper Yom River Basin evaluated using GSMaP-MVK at T+1H, T+6H, T+12H, and T+24H. (a) Predicted discharge during flood periods; (b) scatter plot of predicted vs. observed discharge for the testing dataset.
Panel a displays two line charts comparing observed and predicted river discharge and precipitation over time, with multiple prediction intervals. Panel b contains four scatter plots showing predicted versus observed flow for time horizons of one, six, twelve, and twenty-four hours, each with corresponding correlation coefficients.
Testing dataset results for the Upper Chi River Basin evaluated using GSMaP-MVK at T+1H, T+6H, T+12H, and T+24H. (a) Predicted discharge during flood periods; (b) scatter plot of predicted vs. observed discharge for the testing dataset.
Panel a shows two line graphs comparing observed and predicted river discharge and precipitation over time, with multiple forecast lead times. Panel b features four scatter plots of predicted versus observed flow for different lead times, each displaying a high correlation coefficient.
Figures 6, 7 provide a comparative analysis of the actual and predicted discharge data during the flood events in the internal testing period, alongside the average rainfall obtained from GSMaP-MVK. This visual representation facilitates an examination of the relationship between precipitation within the basin and observed discharge at the downstream station. Theoretically, these two variables could be expected to show a strong correlation, though there may be a noticeable lag between rainfall events and their impact on river discharge. In addition, scatter plots were generated to compare actual and predicted discharge values across various experimental forecasting horizons for both river basins (Figures 6, 7). These plots enable a detailed analysis of model error trends, with points above the diagonal line indicating overpredictions and points below representing underpredictions. Furthermore, the degree of scattering around the diagonal line reflects the accuracy and consistency of the predictions.
In addition, box and whisker plots were constructed to analyze the distribution of actual discharge values with respect to the prediction, as well as the errors between the two, as depicted in Figures 8, 9. These visualizations summarize key statistical characteristics of the data, including the interquartile range (represented by the box), the median (indicated by a horizontal line within the box), and the spread of the data (depicted by the whiskers and any identified outliers). By integrating these plots, the analysis not only identifies systematic biases in the model but also highlights the variability and extremities in the dataset, offering a comprehensive assessment of the model's predictive performance across different scenarios.
Box plots of testing dataset results for the Upper Yom River Basin in dry season (December 1 to April 30) and wet season (May1 to November 30) in (a) the distribution of the observed and predicted discharges over the forecast horizons and (b) the difference between the predicted and actual discharge.
Two box plot charts compare observed and predicted river discharge (m3/s) and associated error for dry and wet periods. Chart (a) shows discharge values for observed and forecasts at multiple lead times. Chart (b) presents forecast errors for the same periods, illustrating increasing discharge variability and errors with longer lead times.
Box plots of testing dataset results for the Upper Chi River Basin in dry season (December 1 to April 30) and wet season (May1 to November 30) in (a) the distribution of the observed and predicted discharges over the forecast horizons and (b) the difference between the predicted and actual discharge.
Two box plot charts compare observed and predicted river discharge (m³/s) and associated error for dry and wet periods. Chart (a) shows discharge values for observed and forecasts at multiple lead times. Chart (b) presents forecast errors for the same periods. Both charts indicate greater variability during wet conditions. Legends and box plot element keys are included for interpretation.
In the Upper Yom River Basin, the considerable average rainfall during 2020 was clearly correlated with a pronounced peak in the discharge data at a subsequent time lag, as illustrated in Figure 6. This lag reflects the time required for the rainfall to flow through the basin and affects the downstream discharge station. During the early rainy season (May to June), due to high infiltration rates, rainfall in the catchment area contributes minimally to runoff at the outlet. In a major flood event in the Upper Yom River Basin during July 2020 (Figure 6a), associated with Tropical Storm Higos, which caused severe flooding in northern Thailand. The predictions demonstrated good agreement with the observed flood hydrograph, particularly at short lead times of 1–12 h. At these horizons, the model closely reproduced the flood hydrograph, accurately capturing both the timing and magnitude of the peak flow. However, as the prediction horizon extended to 24 h, model performance deteriorated, evidenced by increased errors and a delayed peak discharge relative to the observations. In addition, the scatter plot in Figure 6b revealed a clustering of data points around the diagonal line, with some deviations, especially for longer forecasting horizons. The errors remained relatively small for lower discharge values but increased substantially for higher discharges. Both underprediction and overprediction were observed, with overprediction being slightly more prevalent for all forecasting horizons. This trend was further emphasized in Figure 8b, depicting the difference between the predicted and actual discharge. The mean and median differences were positive for all modeled periods (both wet and dry), suggesting overprediction. Error variations, as evidenced by box size, gradually increased with longer forecasting horizons. Furthermore, the box sizes for the wet season were relatively larger than for the dry season, implying greater variability in the former, with some evidence of underprediction. Figure 8a visually represents the distribution of the observed and predicted discharges across forecasting horizons, categorized into dry and wet seasons. In the dry season, the box plots were compact and centered near zero, with short whiskers, indicating minimal errors and well-performed model predictions. Conversely, during the wet season, the box plots expanded, reflecting greater variability in discharge values. The longer whiskers suggested a wider range of errors, accompanied by outliers that underscored the challenges for the model in capturing extreme events or abrupt flow changes. Notably, the box plots for the predicted values closely aligned with those of the observed values, indicating the model's robust performance, even under the dynamic conditions of the wet season.
The Upper Chi River Basin, which experienced comparatively lower precipitation levels, as shown in Figure 7, had a weaker correlation between rainfall and discharge, attributable to the basin's characteristics, including its extensive river system and large catchment areas. Runoff from the upstream areas of the main river and its tributaries would reach the outlet almost simultaneously. As a result, even when rainfall in the tributary catchments was modest, the runoff at the outlet remained relatively high, particularly during the late rainy season. The scatter plots demonstrated an even closer clustering of data points along the diagonal line, particularly for lower discharge values. Deviations from the diagonal generally indicate model underprediction across all forecasting horizons. In Figure 9b supported this observation, showing consistently negative values for the difference between predicted and actual discharge. The relatively similar spread of the box plots across forecasting horizons suggests that the overall distribution of prediction errors does not exhibit a pronounced change with increasing lead time. This may be attributed to the discharge pattern, characterized by a few peaks occurring within a short timeframe. Since most of the training data consisted of relatively low and stable discharge values, the model had a tendency to underpredict higher discharge events. The distributions of observed and predicted discharge, shown in Figure 9a, had similar patterns to those observed in the Upper Yom River Basin.
To further assess the performance of the proposed model across temporal scales and seasonal conditions, Figure 10 presents the average monthly RMSE of discharge predictions at lead times of T + 1H, T + 6H, T + 12H, and T + 24H during the testing period (2019–2022). Panels (a) and (b) show results for the Upper Yom and Upper Chi River basins, respectively. During the dry season, when discharge levels are low, prediction errors are comparable across all lead times in both basins. In contrast, errors increase during the wet season, particularly in months characterized by a wide range of discharge variability. In the Upper Yom River basin, discharge in July ranges from 0 to 1,252 m3/s and exhibits pronounced hourly fluctuations, which coincides with elevated prediction errors at longer lead times. A similar pattern is observed in the Upper Chi River Basin in September. The higher errors are associated with pronounced hourly discharge variability, which poses challenges for the model in accurately capturing rapid flow dynamics compared with periods of more stable discharge conditions.
The average monthly RMSE of discharge forecasts at lead time T + 1H, T + 6H, T + 12H, and T + 24H during the testing period (2019–2022) (a) results for Upper Yom River Basin (b) results for Upper Chi River Basin.
Panel (a) and panel (b) present line graphs comparing root mean square error (RMSE) of predicted discharge at different forecast horizons (T+1H, T+6H, T+12H, T+24H) by month, overlaid with blue and orange bar charts showing hourly discharge ranges for wet and dry seasons, respectively, at two stations for the period 2019 to 2022.External testing dataset results
After training and fine-tuning our models' hyperparameters on the GSMaP-MVK validation set, we collected additional datasets as external test sets. An external test set is a crucial component in validating the performance of deep learning models as it helps in preventing overfitting, assessing generalization, aiding model selection, preventing data leakage, and providing a realistic prediction of real-world performance. By using an external test set, we could ensure that our model was robust, reliable, and ready for deployment.
In this work, we considered recent 2-year data obtained from the downstream station-based discharge data and the satellite-based precipitation data as our external test datasets. Specifically, we collected GSMaP-MVK and GSMaP-NOW data from 2022 to 2024. We prioritized GSMaP-NOW because it offers more up-to-date data with only a 30-min time lag compared to the 3-day lag in GSMaP-MVK. If our model (trained on GSMaP-MVK) demonstrated satisfactory performance on GSMaP-NOW, it would suggest high potential for implementation in a real-time early warning system. The ability to accurately predict discharge based on near-real-time precipitation data is essential for timely flood alerts. It should be noted that although the use of GSMaP-NOW enables relatively near real-time precipitation data within the proposed framework, operational hydrological forecasting systems may still face data latency issues. In practice, both discharge and precipitation data can be affected by reporting delays, data gaps, and quality control procedures, which may result in the unavailability of observations for the most recent hours. This operational limitation is not explicitly addressed in the current study. Future work should evaluate the proposed framework under fully operational conditions and explore strategies to mitigate potential data latency issues. Table 4 summarizes the performance of the model on the external test datasets calculated from GSMaP-MVK and GSMaP-NOW for both study areas.
Performance indices obtained from prediction results of external test dataset.
Performance indices
Upper Yom river basin
Upper Chi river basin
T + 1H
T + 6H
T + 12H
T + 24H
T + 1H
T + 6H
T + 12H
T + 24H
GSMaP-MVK
R
0.9919
0.9791
0.9514
0.8736
0.9989
0.9986
0.9980
0.9957
NSE
0.9826
0.9558
0.8981
0.7459
0.9769
0.9761
0.9757
0.9700
MAE (m3/s)
6.531
9.720
13.881
20.834
11.369
11.360
11.289
11.688
RMSE (m3/s)
13.024
20.009
29.921
46.915
15.233
15.490
15.609
17.365
GSMaP-NOW
R
0.9925
0.9784
0.9476
0.8615
0.9988
0.9985
0.9980
0.9956
NSE
0.9817
0.9510
0.8861
0.7170
0.9767
0.9760
0.9753
0.9697
MAE(m3/s)
6.561
10.068
14.552
21.786
11.363
11.298
11.383
11.676
RMSE(m3/s)
13.332
21.014
31.607
49.493
15.303
15.522
15.764
17.436
Based on the comparison of the evaluation metrics using GSMaP-MVK on the internal test set (Table 3) and the external test set (Table 4), the model performed consistently, with only slight variations between the two datasets. These minor differences suggested no evidence of overfitting, with the variation perhaps due to the different discharge patterns observed during the testing time periods. We further compared the model's performance based on the GSMaP-MVK and GSMaP-NOW datasets, with the performance differences between them being minimal, indicating that the model trained on one dataset was transferable to another dataset. The ability to apply a model trained on GSMaP-MVK to GSMaP-NOW is particularly advantageous for real-world applications. Notably, GSMaP-NOW data has a shorter lag time, making it more suitable for timely decision making and operational use.
Figures 11, 12 present a comparative analysis of actual and predicted discharge during flood periods in the external test dataset for the Upper Yom River Basin, coupled with the average rainfall depth obtained from GSMaP-MVK and GSMaP-NOW, respectively. The model accurately reproduced both peak flood discharges and the timing of flood events across all forecasting lead times. Predicted discharge results for the entire external testing period using GSMaP-MVK and GSMaP-NOW are shown in Supplementary Figures 3, 4, respectively. The model's performances were essentially consistent across datasets, despite slight differences in the rainfall patterns. The consistency of the model performance suggested that our proposed framework was robust and capable of handling input variability. The model effectively captured the major trends, delivering preferable final predictions even in the presence of outliers.
External test dataset for the Upper Yom River Basin evaluated using GSMaP-MVK at T + 1H, T + 6H, T + 12H, and T + 24H. (a) Predicted discharge during flood periods; (b) scatter plot of predicted vs. observed discharge for the external test dataset.
Panel (a) contains two line graphs comparing observed and predicted river discharge over time with corresponding precipitation, while panel (b) shows four scatter plots comparing observed and predicted flow at different forecast lead times, each labeled with correlation coefficients.
External test dataset for the Upper Yom River Basin evaluated using GSMaP-NOW at T + 1H, T + 6H, T + 12H, and T + 24H. (a) Predicted discharge during flood periods; (b) scatter plot of predicted vs. observed discharge for the external test dataset.
Figure contains two main panels labeled (a) and (b). Panel (a) displays two line graphs comparing observed and predicted river discharge against precipitation from August to October, differentiated by temporal prediction intervals. Panel (b) shows four scatterplots, each relating predicted and observed flow for forecast horizons of one, six, twelve, and twenty-four hours, with corresponding correlation coefficients decreasing as forecast interval increases.
Figures 13, 14 provide a similar comparative analysis for the Upper Chi River Basin, while Supplementary Figures 5, 6 show results for the entire external test period using GSMaP-MVK and GSMaP-NOW, respectively. These results reinforce the model's generalizability and its ability to maintain stable performance across different river basins and input data sources. The plots demonstrate that model performance remains largely consistent, despite minor variations in rainfall patterns. Overall, the precipitation trends for both data sources were comparable, with a notable peak observed in GSMaP-NOW during the wet season of 2023 (Supplementary Figure 6). This observation was corroborated by the mass curves, which indicated that the accumulated precipitation for both datasets was similar, with a slight increase in GSMaP-NOW, particularly during this peak period. Based on these results, our proposed model demonstrated robustness and the ability to adapt to subtle variations in the input data. This adaptability is crucial for real-world applications, particularly when the model is deployed as a flooding early warning system. In dynamic and often unpredictable environments, such as those involving fluctuating rainfall patterns and varying hydrological conditions, the model's flexibility ensures reliable performance, making it a strong candidate for operational use in flood prediction and management.
External test dataset for the Upper Chi River Basin evaluated using GSMaP-MVK at T + 1H, T + 6H, T + 12H, and T + 24H. (a) Predicted discharge during flood periods; (b) scatter plot of predicted vs. observed discharge for the external test dataset.
Figure with two panels labeled a and b. Panel a contains two time series charts displaying observed and predicted river discharge against precipitation from September to November, with a sharp peak in October and multiple lines for prediction intervals. Panel b contains four scatter plots comparing predicted and observed river flow at forecast horizons of one, six, twelve, and twenty-four hours, demonstrating strong linear correlation and high R values above 0.99.
External test dataset for the Upper Chi River Basin evaluated using GSMaP-NOW at T + 1H, T + 6H, T + 12H, and T + 24H. (a) Predicted discharge during flood periods; (b) scatter plot of predicted vs. observed discharge for the external test dataset.
Panel (a) contains two time series plots showing river discharge in cubic meters per second overlaid with precipitation data for two different periods, with multiple predicted and observed discharge lines and a red precipitation bar chart. Panel (b) features four scatter plots comparing predicted versus observed flow for lead times of one, six, twelve, and twenty-four hours, each with a line indicating perfect agreement and R values above 0.99.
The use of satellite-based data has proven to be promising for predicting hydrological conditions in data-scarce areas, such as the regions in our study. Our model delivered reasonably good performance, even for extended forecasting horizons under realistic data constraints, while acknowledging that completely ungauged forecasting remains an important topic for future research. In addition, further refinements are necessary to optimize the model's effectiveness for long-term predictions, especially in early warning systems. Several strategies could be considered to enhance the model's performance for extended forecasts. Fine-tuning model parameters or integrating more advanced algorithms may enhance accuracy, especially in predicting extreme events. Additionally, incorporating variables, such as rainfall intensity, antecedent soil moisture conditions, or land use changes, could significantly improve the model's predictive capabilities. For real-world deployment, further testing with a larger and more diverse dataset, especially one that includes a broader range of extreme events, would offer a more comprehensive evaluation of the model's robustness and reliability. Moreover, extending the proposed framework to basins in different geographical regions may introduce additional challenges related to basin-scale variability, which remain to be explored in future work. To address these challenges, adaptive input representations—such as basin-normalized spatial windows, multi-resolution pooling strategies, or attention-based spatial weighting mechanisms—could be incorporated to better capture scale-dependent precipitation patterns. Such developments may enhance model transferability across basins of varying sizes and physiographic characteristics, thereby improving its suitability for operational early warning applications where accurate prediction of extreme hydrological events is critical for timely and effective decision-making.
In this study, all input data, including GSMaP precipitation products and discharge observations, are obtained from publicly available sources and responsible organizations. These datasets do not contain personal, proprietary, or sensitive information. Data processing and modeling procedures are documented to ensure transparency and reproducibility. However, the application of the proposed model for flood forecasting requires careful consideration of data limitations and potential bias, as historical precipitation and discharge datasets may reflect legacy effects of past water management practices and uneven spatial and temporal monitoring coverage. These issues are particularly relevant in flood early warning applications, where model outputs may directly support risk communication and emergency response. Therefore, the results should be interpreted as decision-support information rather than definitive predictions, with explicit recognition of uncertainty and the continued role of expert judgment in responsible flood risk management.
Conclusion
The proposed network architecture effectively integrated both spatial and temporal information, leveraging the strengths of CNNs for spatial feature extraction and LSTMs for temporal dependency modeling. This approach demonstrated potential for satellite data analysis and time series forecasting in hydrological contexts. Internal and external validation using the GSMaP-MVK and GSMaP-NOW datasets confirmed the model's reliability and generalization across different data scenarios, especially for real-time conditions. The integration of advanced deep learning techniques and satellite data offers a scalable and practical solution for enhancing discharge predictions in data-limited regions. Our proposed framework should not only improve hydrological forecast accuracy but should also support proactive water resource management, particularly in catchments where upstream ground-based measurements are sparse or incomplete. Potential future research directions include expanding the model by incorporating additional environmental variables, exploring more sophisticated neural network architectures, and testing the approach across diverse hydrological systems to validate and extend its broader applicability.
Data availability statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Author contributions
PW: Supervision, Software, Data curation, Writing – original draft, Writing – review & editing, Conceptualization, Methodology, Validation. KY: Writing – original draft, Visualization, Formal analysis, Methodology. WT: Funding acquisition, Investigation, Supervision, Data curation, Writing – original draft, Project administration, Conceptualization, Methodology.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frwa.2026.1753598/full#supplementary-material
ReferencesChangF.-J.ChangL.-C.ChenJ.-F. (2023). Artificial intelligence techniques in hydrology and water resources management. Water15:1846. doi: 10.3390/w15101846ChenminN.MarsaniM. F.Pei ShanF. (2024). Flood prediction based on feature selection and a hybrid deep learning network. J. Water Clim. Change15, 1245–1261. doi: 10.2166/wcc.2024.559FengJ.ShaH.DingY.YanL.YuZ. (2022). “Graph convolution based spatial-temporal attention LSTM model for flood forecasting,” in Paper Presented at the 2022 International Joint Conference on Neural Networks (IJCNN) (Padua: IEEE). doi: 10.1109/IJCNN55064.2022.9892371HashemiM. (2019). Enlarging smaller images before inputting into convolutional neural network: zero-padding vs. interpolation. J. Big Data6, 1–13. doi: 10.1186/s40537-019-0263-7HuF.YangQ.YangJ.LuoZ.ShaoJ.WangG.. (2024). Incorporating multiple grid-based data in CNN-LSTM hybrid model for daily runoff prediction in the source region of the Yellow River Basin. J. Hydrol. Reg. Stud.51:101652. doi: 10.1016/j.ejrh.2023.101652KratzertF.KlotzD.BrennerC.SchulzK.HerrneggerM. (2018). Rainfall-runoff modelling using long short-term memory (LSTM) networks. Hydrol. Earth Syst. Sci.22, 6005–6022. doi: 10.5194/hess-22-6005-2018KubotaT.AonashiK.UshioT.ShigeS.TakayabuY. N.KachiM.. (2020). “Global satellite mapping of precipitation (GSMaP) products in the GPM era,” in Satellite Precipitation Measurement (Cham: Springer), 355–373. doi: 10.1007/978-3-030-24568-9_20KumarV.AzamathullaH. M.SharmaK. V.MehtaD. J.MaharajK. T. (2023). The state of the art in deep learning applications, challenges, and future prospects: a comprehensive review of flood forecasting and management. Sustainability15:10543. doi: 10.3390/su151310543LeeJ.LeeJ. E.KimN. W. (2020). Estimation of hourly flood hydrograph from daily flows using artificial neural network and flow disaggregation technique. Water13:30. doi: 10.3390/w13010030LiP.ZhangJ.KrebsP. (2022). Prediction of flow based on a CNN-LSTM combined deep learning approach. Water14:993. doi: 10.3390/w14060993LiX.XuW.RenM.JiangY.FuG. (2022). Hybrid CNN-LSTM models for river flow prediction. Water Supply22, 4902–4919. doi: 10.2166/ws.2022.170MalikH.FengJ.ShaoP.AbduljabbarZ. A. (2024). Improving flood forecasting using time-distributed CNN-LSTM model: a time-distributed spatiotemporal method. Earth Sci. Inf.17, 1–20. doi: 10.1007/s12145-024-01354-ySenent-AparicioJ.Jimeno-SáezP.Bueno-CrespoA.Pérez-SánchezJ.Pulido-VelázquezD. (2019). Coupling machine-learning techniques with SWAT model for instantaneous peak flow prediction. Biosyst. Eng.177, 67–77. doi: 10.1016/j.biosystemseng.2018.04.022ShenC.ChenX.LaloyE. (2021). Broadening the use of machine learning in hydrology. Front Water3:681023. doi: 10.3389/frwa.2021.681023SitM.DemirayB. Z.XiangZ.EwingG. J.SermetY.DemirI.. (2020). A comprehensive review of deep learning applications in hydrology and water resources. Water Sci. Technol.82, 2635–2670. doi: 10.2166/wst.2020.36933341760SongC. M. (2022). Data construction methodology for convolution neural network based daily runoff prediction and assessment of its applicability. J. Hydrol.605:127324. doi: 10.1016/j.jhydrol.2021.127324ThaisiamW.SaeloW.WongchaisuwatP. (2022). Enhancing a multi-step discharge prediction with deep learning and a response time parameter. Water14:2898. doi: 10.3390/w14182898ThaisiamW.YomwilaiK.WongchaisuwatP. (2024). Utilizing sequential modeling in collaborative method for flood forecasting. J. Hydrol.636:131290. doi: 10.1016/j.jhydrol.2024.131290TongalH.BooijM. J. (2018). Simulation and forecasting of streamflows using machine learning models coupled with base flow separation. J. Hydrol.564, 266–282. doi: 10.1016/j.jhydrol.2018.07.004WangY.LiuJ.XuL.YuF.ZhangS. (2023). Streamflow simulation with high-resolution WRF input variables based on the CNN-LSTM hybrid model and gamma test. Water15:1422. doi: 10.3390/w15071422WaqasM.HumphriesU. W. (2024). A critical review of RNN and LSTM variants in hydrological time series predictions. MethodsX13:102946. doi: 10.1016/j.mex.2024.10294639324077WMO (2022). Early Warning System. Geneva: WMO.WMO (2024). “State of the climate in Asia 2023 (WMO-No.1350),” in Paper Presented at the 80th Session of the Commission (ESCAP) (Bangkok).WuW.EmertonR.DuanQ.WoodA. W.WetterhallF.RobertsonD. E.. (2020). Ensemble flood forecasting: current status and future opportunities. Water7:e1432. doi: 10.1002/wat2.1432YaoZ.WangZ.WangD.WuJ.ChenL. (2023). An ensemble CNN-LSTM and GRU adaptive weighting model based improved sparrow search algorithm for predicting runoff using historical meteorological and runoff data as input. J. Hydrol.625:129977. doi: 10.1016/j.jhydrol.2023.129977ZanchettaA. D.CoulibalyP.FortinV. (2022). Forecasting high-flow discharges in a flashy catchment using multiple precipitation estimates as predictors in machine learning models. Hydrology9:216. doi: 10.3390/hydrology9120216ZhangJ.YanH. (2023). A long short-term components neural network model with data augmentation for daily runoff forecasting. J. Hydrol.617:128853. doi: 10.1016/j.jhydrol.2022.128853
Edited by: Alyssa Dausman, The Water Institute of the Gulf, United States
Reviewed by: Yuan Yang, University of California, San Diego, United States