JEERESD

Volume 4 · Issue 4 · 2026 · Pages 52–80 · Published online 1 October 2026

Original Article · English

Characterization and Forecasting of Future Drought in the Gambia River Basin at Kédougou Using Standardized Indices, Machine Learning, and Deep Neural Networks

Ibrahima Faye · Cheikh Faye

Department of Geography, Faculty of Science and Technology, Assane Seck University of Ziguinchor, Laboratory of Geomatics and Environment

DOI10.68305/jeeresd.v4i4.005

01Abstract

Background: Drought is one of the most complex and damaging hydroclimatic hazards, particularly in West Africa where water resources are highly sensitive to climate variability and change. This study analyzes the future evolution of meteorological, climatic, and hydrological droughts in the Gambia River basin controlled at the Kédougou gauging station in southeastern Senegal. Methods: The approach combines precipitation, potential evapotranspiration (PET), temperature, and simulated streamflow, and computes the SPI, SPEI, and SSI at 1-, 3-, 6-, and 12-month timescales. Results: Results reveal a major contrast between the precipitation signal and thermo-hydrological responses. Under SSP5-8.5, mean temperature increases from 29.67 °C in 2021–2040 to 33.99 °C in 2081–2100, annual PET rises from 1,574 to 1,770 mm, and mean streamflow decreases from 49.02 to 12.45 m³/s. Mean SSI-3 simultaneously deteriorates from −1.21 to −1.94, while the frequency of drought months increases from 67.5% to 91.7% and that of extreme drought months from 35.4% to 84.2%. Under SSP1-2.6, drying remains substantial but is less monotonic. Eight forecasting models are compared at 1-, 3-, and 6-month lead times: Random Forest, ExtraTrees, XGBoost, LightGBM, SVR, MLP/ANN, LSTM, and GRU. LightGBM performs best at one month (NSE = 0.867 under SSP1-2.6 and 0.718 under SSP5-8.5), whereas LSTM/GRU become competitive at longer lead times. Conclusion: The study highlights a risk of chronic hydrological drought by the end of the century, particularly under high emissions, and demonstrates the value of multimodel forecasting for early warning and adaptive water-resources management.

KeywordsHydrological drought; SSI; SPEI; SPI; Climate change; Gambia River

02Author contributions

Ibrahima Faye

Conceptualization; Methodology; Software; Formal Analysis; Investigation; Data Curation; Visualization; Writing - Original Draft

Cheikh Faye

Conceptualization; Methodology; Investigation; Data Curation; Resources; Supervision; Project Administration; Validation; Writing - Review & Editing

03What this study adds

This study makes several original contributions to the understanding of future drought dynamics in West Africa: 1. Integrated Multi-Index Drought Assessment This study provides a comprehensive characterization of meteorological, climatic, and hydrological drought in the Gambia River basin at Kédougou by jointly analyzing SPI, SPEI, and SSI. It demonstrates that precipitation-based indices alone (SPI) substantially underestimate future drought severity, as they fail to capture the amplifying effects of rising temperature and evaporative demand. This integrated approach reveals that while SPI remains relatively moderate, SPEI and SSI show severe deterioration, particularly under high-emission scenarios. 2. Contrasting SSP Scenario Trajectories The study quantifies the divergence between low-emission (SSP1-2.6) and high-emission (SSP5-8.5) pathways for the upper Gambia River basin. Under SSP5-8.5, mean temperature increases by 4.32 °C, PET by 12.4%, and streamflow collapses by 74.6% by 2081–2100, with 91.7% of months classified as drought and 84.2% as extreme drought (SSI-3 ≤ −2). These findings indicate a potential transition from episodic drought toward chronic hydrological drought, providing critical evidence for long-term water-resource planning. 3. Multi-Model Artificial-Intelligence Forecasting This study is among the first to evaluate eight machine-learning and deep-learning algorithms (Random Forest, ExtraTrees, XGBoost, LightGBM, SVR, MLP/ANN, LSTM, and GRU) for forecasting SSI-3 at 1-, 3-, and 6-month lead times in a West African context. It demonstrates that LightGBM excels at short-term forecasting (NSE = 0.867 under SSP1-2.6), while LSTM and GRU become competitive at longer lead times. However, predictive skill declines sharply at 3–6 months under SSP5-8.5, highlighting the challenge of forecasting in a highly non-stationary climate. 4. Identification of Controlling Predictors The study identifies the dominant hydroclimatic variables controlling forecast skill: short-term forecasts depend primarily on hydrological persistence (lagged SSI and streamflow), while longer lead times require precipitation and PET, emphasizing the importance of continuous hydrometric monitoring for operational drought early-warning systems. 5. Policy-Relevant Implications By projecting quasi-chronic hydrological drought by the end of the century under high emissions, this study provides actionable evidence for water-resource managers and policymakers in Senegal and the wider Gambia River basin. It underscores the need for strengthening hydroclimatic monitoring, conserving recharge areas, improving water-use efficiency, and developing transboundary information-sharing mechanisms to enhance adaptive capacity.

04Introduction

Drought is one of the most complex and damaging hydroclimatic hazards because of its slow onset, potentially long duration, large spatial extent, and capacity to affect simultaneously water resources, agricultural production, ecosystems, energy systems, and socioeconomic activities. Unlike rapidly developing hazards, drought generally results from the progressive accumulation of hydroclimatic anomalies and can propagate through different components of the hydrological cycle (Mishra & Singh, 2010; Wang et al., 2022). A persistent precipitation deficit may first produce meteorological drought before affecting soil moisture, groundwater recharge, river discharge, reservoir storage, and ultimately water availability. Consequently, drought cannot be adequately characterized through precipitation anomalies alone, particularly in regions where rising temperatures and increasing atmospheric evaporative demand may amplify water deficits (Ahmadalipour et al., 2017; Vicente-Serrano et al., 2010). This issue is particularly important in West Africa, where water resources are highly sensitive to climate variability and change. The region experienced severe and persistent drought episodes during the second half of the twentieth century, accompanied by substantial changes in rainfall and river-flow regimes. Although rainfall recovery has been observed in several parts of the Sahel since the late 1990s, strong interannual and spatial variability persists, and hydrological responses do not necessarily follow precipitation trends proportionally (Traore et al., 2014; Rameshwaran et al., 2021). Future climate change could further modify these relationships through changes in precipitation distribution, temperature, evapotranspiration, rainfall extremes, and land-surface processes. Recent climate projections therefore raise important concerns regarding the future reliability of surface-water resources in West African river basins (Bodian et al., 2018; Diop et al., 2025). The Gambia River basin constitutes a particularly relevant system for investigating these changes. Originating in the Fouta Djallon highlands in Guinea and extending through Senegal and The Gambia, the basin supports agriculture, livestock production, domestic water supply, ecosystems, biodiversity conservation, and numerous socioeconomic activities. Its hydrological functioning is strongly controlled by the West African monsoon and by the marked seasonal concentration of precipitation (Dione, 1996; Sow, 2007). Historical analyses have demonstrated significant variations and shifts in rainfall and runoff within the basin, highlighting its sensitivity to regional hydroclimatic fluctuations (Traore et al., 2014; Faye & Mendy, 2018). More recent studies have also shown that climate variability, land-use dynamics, and changes in rainfall–runoff relationships may substantially affect the basin's water resources (Kah et al., 2025). This vulnerability could intensify during the twenty-first century. Climate-model simulations indicate potentially important changes in West African river flows under global warming (Rameshwaran et al., 2021). In the Senegal and Gambia river basins, projected climate changes have already been associated with substantial modifications of future streamflow (Bodian et al., 2018). More specifically, recent assessments of the Gambia River basin indicate that changes in precipitation, temperature, and evapotranspiration may translate into significant changes in water availability and hydrological regimes (Séne et al., 2024). Such projections emphasize the need to assess drought not merely as a meteorological phenomenon but as a process propagating from climatic forcing to hydrological response. Several standardized indices have been developed to characterize these complementary dimensions of drought. The Standardized Precipitation Index (SPI) is widely used to quantify meteorological drought through precipitation anomalies accumulated over different timescales. However, precipitation-based indices may underestimate drought intensification under warming conditions because they do not explicitly account for increasing atmospheric evaporative demand. The Standardized Precipitation Evapotranspiration Index (SPEI) addresses this limitation by incorporating the climatic water balance between precipitation and potential evapotranspiration and is therefore particularly suited to drought assessment in a warming climate (Vicente-Serrano et al., 2010). At the hydrological level, standardized runoff or streamflow indices provide direct information on anomalies in river discharge and water availability (Shukla & Wood, 2008). The Standardized Streamflow Index (SSI), in particular, enables hydrological drought to be characterized consistently across different temporal scales (Vicente-Serrano et al., 2012). Joint analysis of SPI, SPEI, and SSI can therefore provide a more comprehensive understanding of drought propagation from atmospheric deficits to river-flow responses. The importance of such an integrated framework increases under future climate conditions because drought propagation is nonlinear and influenced by multiple interacting controls. Temperature rise can intensify evapotranspiration and modify the conversion of rainfall into runoff, meaning that relatively moderate precipitation changes may generate disproportionately large hydrological responses (Ahmadalipour et al., 2017; Wang et al., 2022). Land-use and land-cover changes can further alter infiltration, runoff generation, and basin storage, thereby modifying hydrological drought characteristics (Kah et al., 2025; Liu et al., 2025). Moreover, climate projections themselves contain uncertainties related to emission pathways, climate-model structure, and the representation of regional hydroclimatic processes (Abdulai & Chung, 2019; Song et al., 2021). Assessing drought under contrasting Shared Socioeconomic Pathways is consequently important for distinguishing the implications of lower- and higher-emission futures. The SSP framework provides internally consistent trajectories of socioeconomic development and greenhouse-gas forcing (Riahi et al., 2017), while ScenarioMIP provides the coordinated climate-model framework used to investigate these alternative futures (O'Neill et al., 2016). Beyond drought characterization, reliable forecasting is essential for translating climate information into operational early-warning and water-management decisions. Traditional statistical and process-based hydrological approaches remain fundamental, but the increasing availability of hydroclimatic datasets has stimulated the development of machine-learning and deep-learning methods capable of representing complex nonlinear relationships. Random Forest, for example, provides a flexible ensemble-learning framework capable of handling nonlinear interactions among numerous predictors (Breiman, 2001). Recurrent neural-network architectures such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models are particularly promising because they can learn temporal dependencies and hydrological memory from sequential data. Recent reviews demonstrate the rapidly increasing application of deep-learning techniques to streamflow, hydrological, and drought forecasting (Zhao et al., 2024). Nevertheless, predictive performance depends strongly on forecast horizon, input-variable selection, hydroclimatic regime, and the stationarity of relationships between predictors and drought indicators. Despite these methodological advances, relatively few studies have jointly investigated future meteorological, climatic, and hydrological drought while simultaneously evaluating the predictability of hydrological drought at several lead times in the upper Gambia River basin. Existing research has substantially improved understanding of historical hydroclimatic variability, rainfall–runoff relationships, future streamflow, and climate-change impacts in the basin (Faty et al., 2018; Faty et al., 2023; Séne et al., 2024; Kah et al., 2025), but an integrated assessment combining multiple standardized drought indices with machine-learning and deep-learning forecasting remains necessary. Such an approach is especially relevant at Kédougou, where the upstream catchment provides an important hydrological control point for understanding climate-induced changes in the continental Gambia River system. Accordingly, this study aims to characterize and forecast future drought in the Gambia River basin at the Kédougou gauging station under contrasting climate-change scenarios. Specifically, it seeks to: (i) quantify future meteorological, climatic, and hydrological drought using SPI, SPEI, and SSI at multiple timescales; (ii) examine the evolution of drought intensity, frequency, duration, and severity under SSP1-2.6 and SSP5-8.5 relative to the 1985–2014 reference climatology; (iii) evaluate eight machine-learning and deep-learning models—Random Forest, ExtraTrees, XGBoost, LightGBM, SVR, MLP/ANN, LSTM, and GRU—for forecasting SSI-3 at 1-, 3-, and 6-month lead times; and (iv) identify the principal hydroclimatic predictors controlling forecast skill. By integrating climate-change projections, standardized drought indicators, and artificial-intelligence-based forecasting, the study seeks to provide a scientifically robust framework for drought early warning and adaptive water-resources management in the Gambia River basin.

05Materials and methods

2.1 Study Area The Gambia River basin is one of the major river systems in West Africa. It covers approximately 77,100 km² across three countries: Guinea (15.25% of the basin area), Senegal (72.30%), and The Gambia (12.45%). The Gambia River rises in the Fouta Djallon highlands in Guinea, at an elevation of approximately 1,150 m near Labé, before flowing nearly 1,180 km to its mouth in the Atlantic Ocean in The Gambia (Kah et al., 2025). It comprises two main sections: a continental reach and a maritime reach (Dione, 1996; Sow, 2007; Faye & Mendy, 2018). The continental reach receives numerous tributaries on its left bank, including the Diagueri, Niokolo-Koba, Niéri-Ko, and Sandougou, and on its right bank, notably the Thiokoye, Diarha, and Koulountou (Figure 1). Its hydrological regime is tropical Sudano-Guinean and is strongly influenced by the West African monsoon, with a marked alternation between a long dry season (November to May) and a rainy season (June to October). Streamflow is mainly rainfall-driven, with precipitation concentrated between July and September, when the principal floods occur (Séne et al., 2024). The Kédougou gauging station, located in southeastern Senegal, controls a catchment of approximately 7,550 km² and therefore represents an important part of the upper Gambia River basin. The area is characterized by relatively rugged topography dominated by the foothills of the Fouta Djallon, Precambrian basement geological formations, and vegetation consisting mainly of wooded savannas and open forests. Annual rainfall, generally ranging from 1,000 to 1,400 mm, is among the highest in Senegal and explains the substantial streamflow recorded at the station. Discharge shows strong seasonal variability, with pronounced low flows during the dry season and rapid floods during the wet season. Because of its upstream position and the quality of its hydrometric records, the Kédougou station is a reference site for investigating hydrological variability, climate-change impacts on water resources, rainfall–runoff modeling, and river-flood risk in the Gambia River basin. Figure 1: Gambia River basin at the Kédougou station 2.2 Hydroclimatic Data and Climate Scenarios The methodological framework combines climatic, hydrological, and climate-projection datasets to characterize the evolution of drought and assess its predictability in the Gambia River basin at the Kédougou gauging station. Four principal hydroclimatic variables are considered: precipitation (P), potential evapotranspiration (PET), mean air temperature (T), and streamflow (Q). The Kédougou station controls an upstream catchment of approximately 7,550 km² characterized by a tropical Sudano-Guinean hydrological regime strongly influenced by the West African monsoon. The marked seasonality of rainfall and runoff makes this catchment particularly suitable for examining the propagation of meteorological anomalies into hydrological drought. Historical precipitation is derived from the Climate Hazards Group InfraRed Precipitation with Stations (CHIRPS) dataset. CHIRPS combines satellite-based infrared observations with station measurements to generate quasi-global rainfall estimates at a spatial resolution of 0.05°, making it particularly useful in regions characterized by sparse or spatially heterogeneous rain-gauge networks (Funk et al., 2015). Previous applications in the Gambia River basin have demonstrated the potential of satellite rainfall products for hydrological simulation and streamflow reconstruction (Faty et al., 2018; Faty et al., 2023). CHIRPS was therefore selected to provide spatially consistent historical precipitation information over the study catchment. Mean air temperature is obtained from ERA5-Land, the land component of the ERA5 reanalysis. ERA5-Land provides temporally consistent, high-resolution estimates of land-surface and atmospheric variables and has become an important dataset for hydrological and environmental applications (Muñoz-Sabater et al., 2021). Observed streamflow data are obtained from the Kédougou hydrometric station managed by Senegal's Directorate for Water Resources Management and Planning (DGPRE). Potential evapotranspiration is estimated from air temperature using the parsimonious formulation proposed by Oudin et al. (2005), which was specifically developed for rainfall–runoff modelling and offers the advantage of limited meteorological data requirements. Future streamflow deserves particular methodological attention because discharge is not directly provided by the climate-model forcing datasets used in this study. Future Q series are therefore generated through rainfall–runoff simulation using a Random Forest (RF) model. Random Forest is an ensemble learning algorithm based on multiple regression trees and is well suited to modelling nonlinear relationships and complex interactions among hydroclimatic predictors (Breiman, 2001). The RF rainfall–runoff model was calibrated over 1992–1999 and independently validated over 2000–2006. Predictors included precipitation, PET, temperature, and derived hydroclimatic information such as lagged variables, rolling accumulations, and seasonal components. The resulting Nash–Sutcliffe efficiency reached 0.976 during calibration and 0.928 during validation, indicating a high ability to reproduce observed streamflow dynamics and supporting application of the model to generate future discharge series. Future climatic forcing is obtained from the Inter-Sectoral Impact Model Intercomparison Project phase 3b (ISIMIP3b), which provides bias-adjusted climate forcing derived from CMIP6 simulations. Two contrasting pathways are analyzed: SSP1-2.6, representing a low-emission and strong-mitigation future, and SSP5-8.5, representing a high-emission pathway. These scenarios provide a useful framework for evaluating the sensitivity of future water resources to contrasting greenhouse-gas trajectories (O'Neill et al., 2016; Riahi et al., 2017). The period 1985–2014, corresponding to 360 monthly observations, is used exclusively as the reference climatology for standardization. Daily precipitation and PET are aggregated as monthly totals, whereas temperature and streamflow are represented by monthly means. Future conditions are summarized over four successive 20-year periods—2021–2040, 2041–2060, 2061–2080, and 2081–2100—to assess the progressive evolution of hydroclimatic conditions and drought throughout the twenty-first century. 2.3 Drought Indices Three complementary standardized drought indices are calculated to represent different stages of drought development: the Standardized Precipitation Index (SPI), Standardized Precipitation Evapotranspiration Index (SPEI), and Standardized Streamflow Index (SSI). Their combined application enables drought propagation to be investigated from atmospheric forcing to climatic water balance and ultimately river-flow response. SPI quantifies precipitation anomalies accumulated over different temporal scales and is widely used for meteorological drought monitoring because of its standardized and multiscalar nature (McKee et al., 1993). However, SPI depends exclusively on precipitation and therefore does not explicitly represent the effect of rising temperature and atmospheric evaporative demand. SPEI addresses this limitation by standardizing the climatic water balance, P−PET, thereby incorporating the influence of evapotranspiration and making the index particularly relevant for drought assessment under global warming (Vicente-Serrano et al., 2010). SSI applies a comparable standardization framework to streamflow and is used here as the primary indicator of hydrological drought. Streamflow-based standardized indices directly reflect anomalies in surface-water availability and therefore provide information that cannot necessarily be inferred from precipitation alone (Shukla & Wood, 2008; Vicente-Serrano et al., 2012). SPI, SPEI, and SSI are calculated at 1-, 3-, 6-, and 12-month accumulation periods to represent drought processes operating at different temporal scales. Standardization is performed separately for each calendar month using empirical distributions fitted exclusively over the 1985–2014 reference period and subsequently transformed into standard normal scores. Maintaining a fixed historical reference allows future anomalies to be interpreted relative to a common climatological baseline while preserving the strong seasonal structure of the basin. Detailed hydrological drought analysis focuses on SSI-3 because this three-month timescale provides a compromise between short-term responsiveness and hydrological memory. Months with SSI-3 ≤ −1 are classified as drought, SSI-3 ≤ −1.5 as severe drought, and SSI-3 ≤ −2 as extreme drought. Drought events are defined as uninterrupted sequences of months satisfying the drought threshold. For each event, duration, minimum SSI-3, cumulative severity, and mean streamflow are calculated. 2.4 Artificial-Intelligence-Based Forecasting Eight machine-learning and deep-learning algorithms are compared for forecasting SSI-3: Random Forest, ExtraTrees, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Support Vector Regression (SVR), multilayer perceptron/artificial neural network (MLP/ANN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). This multimodel framework was adopted because different algorithms represent nonlinearities and temporal dependencies differently. XGBoost implements scalable gradient-boosted decision trees (Chen & Guestrin, 2016), whereas LightGBM improves computational efficiency through gradient-based one-side sampling and feature bundling (Ke et al., 2017). LSTM networks were specifically designed to retain information over long sequential dependencies and overcome limitations associated with vanishing gradients in conventional recurrent networks (Hochreiter & Schmidhuber, 1997), while GRU provides a gated recurrent architecture capable of representing temporal dependencies with a comparatively parsimonious structure (Cho et al., 2014). Predictors comprise precipitation, PET, temperature, streamflow, SPI-3, SPEI-3, SSI-1, SSI-3, and SSI-6. Temporal memory is incorporated using lags of 0, 1, 2, 3, 6, and 12 months, while LSTM and GRU models use 12-month input sequences. The prediction target is SSI-3 at lead times of 1, 3, and 6 months. To minimize information leakage, data are split strictly chronologically and independently for each climate scenario. Targets up to 2050 are used for training, 2051–2075 for validation, and 2076–2100 for independent testing. Continuous predictive performance is evaluated using RMSE, MAE, R², Pearson correlation coefficient (r), Nash–Sutcliffe efficiency (NSE), and Kling–Gupta efficiency (KGE). Drought detection (SSI-3 ≤ −1) is additionally evaluated using accuracy, precision, recall, and F1 score. This dual evaluation distinguishes the ability to predict continuous drought intensity from the ability to correctly detect drought occurrence. 2.5 Descriptive Statistics and Comparative Analysis Finally, descriptive and comparative analyses are performed separately for each scenario and future period. Mean annual precipitation, PET, temperature, and streamflow are calculated together with mean SPI-3, SPEI-3, and SSI-3. The proportions of months experiencing drought, severe drought, and extreme drought are then quantified. Drought events are characterized by their frequency, mean and maximum duration, minimum index value, and cumulative severity. Comparisons between SSP1-2.6 and SSP5-8.5 are used to identify scenario-dependent changes in drought magnitude and persistence. Finally, predictor-importance estimates from tree-based models are examined to identify the dominant climatic and hydrological controls on SSI-3 forecasting and to determine how the relative importance of short-term hydrological persistence and antecedent climate forcing changes with increasing forecast horizon.

06Results

3.1 Future Hydroclimatic Changes Under SSP1-2.6, annual precipitation does not follow a monotonic trend: it reaches 1,015.5 mm yr⁻¹ in 2021–2040 and 1,111.0 mm yr⁻¹ in 2041–2060, decreases to 925.4 mm yr⁻¹ in 2061–2080, and rises again to 1,101.5 mm yr⁻¹ in 2081–2100 (Table 1). In contrast, PET increases from 1,574.4 to 1,604.0 mm yr⁻¹ and temperature from 29.67 to 30.32 °C. Mean streamflow decreases from 44.41 to 35.21 m³/s between the beginning and end of the century, with a minimum of 33.59 m³/s in 2061–2080. Table 1. Future hydroclimatic statistics by scenario and period. Scenario Period P (mm/yr) PET (mm/yr) T (°C) Mean Q (m³/s) SSP1-2.6 2021-2040 1015.5 1574.4 29.67 44.41 SSP1-2.6 2041-2060 1111.0 1588.0 29.97 46.49 SSP1-2.6 2061-2080 925.4 1602.9 30.30 33.59 SSP1-2.6 2081-2100 1101.5 1604.0 30.32 35.21 SSP5-8.5 2021-2040 1101.6 1574.5 29.67 49.02 SSP5-8.5 2041-2060 943.1 1634.3 30.99 27.57 SSP5-8.5 2061-2080 1063.9 1703.4 32.51 17.92 SSP5-8.5 2081-2100 994.6 1770.3 33.99 12.45 Under SSP5-8.5, the thermal signal is much stronger. Mean temperature rises from 29.67 °C in 2021–2040 to 33.99 °C in 2081–2100 (+4.32 °C), while annual PET increases from 1,574.5 to 1,770.3 mm (+12.4%) (Figure 2). Precipitation varies between approximately 943 and 1,102 mm yr⁻¹ without a comparable monotonic decline. Nevertheless, mean streamflow collapses from 49.02 to 12.45 m³/s, a 74.6% decrease between these two periods. This divergence suggests that future hydrological drought cannot be interpreted from mean precipitation alone: increasing evaporative demand, seasonality, and nonlinear rainfall–runoff behavior play decisive roles. Figure 2. Joint evolution of mean streamflow and temperature across the four future periods. 3.2 Characterization Using SPI-3, SPEI-3, and SSI-3 The contrast between SPI-3 and SPEI-3 is one of the major findings. Under SSP5-8.5, mean SPI-3 remains between −0.39 and −0.11 depending on the period, whereas SPEI-3 reaches −1.59 in both 2041–2060 and 2081–2100 (Table 2). Incorporating PET therefore reveals a much stronger climatic deficit than precipitation alone would suggest. The same mechanism is evident under SSP1-2.6: in 2081–2100, SPI-3 is −0.22 whereas SPEI-3 reaches −1.14. Table 2. Mean indices and frequency of hydrological drought classes. Scenario Period SPI-3 SPEI-3 SSI-3 Drought (%) Severe (%) Extreme (%) SSP1-2.6 2021-2040 -0.39 -1.16 -1.44 74.6 57.9 43.3 SSP1-2.6 2041-2060 -0.06 -1.01 -1.26 68.3 48.3 34.6 SSP1-2.6 2061-2080 -0.62 -1.54 -1.69 86.7 71.2 60.4 SSP1-2.6 2081-2100 -0.22 -1.14 -1.54 80.4 64.2 54.2 SSP5-8.5 2021-2040 -0.14 -0.85 -1.21 67.5 49.2 35.4 SSP5-8.5 2041-2060 -0.39 -1.59 -1.75 87.5 77.1 64.6 SSP5-8.5 2061-2080 -0.11 -1.42 -1.84 88.3 81.2 74.2 SSP5-8.5 2081-2100 -0.35 -1.59 -1.94 91.7 88.3 84.2 The hydrological response is even more severe. Under SSP1-2.6, mean SSI-3 is already below −1 in all periods: −1.44 in 2021–2040, −1.26 in 2041–2060, −1.69 in 2061–2080, and −1.54 in 2081–2100. The 2061–2080 period is the most unfavorable, with 86.7% of months in drought, 71.2% in severe drought, and 60.4% in extreme drought. The slight improvement in 2081–2100 does not return the system to historical conditions (Figure 3 and Figure 4). Under SSP5-8.5, deterioration is progressive and very pronounced. Mean SSI-3 decreases from −1.21 in 2021–2040 to −1.75 in 2041–2060, −1.84 in 2061–2080, and −1.94 in 2081–2100. The frequency of drought months increases from 67.5% to 91.7%. Even more strikingly, extreme drought months rise from 35.4% to 84.2%, meaning that more than four out of five months in the final period fall below the SSI-3 ≤ −2 threshold. The high-emissions pathway therefore leads to conditions approaching chronic hydrological drought relative to the 1985–2014 climatology. Figure 3. Evolution of mean SPI-3, SPEI-3, and SSI-3 by period and scenario. Figure 4. Proportion of months affected by hydrological drought (SSI-3 ≤ −1). 3.3 Duration and Severity of Drought Events The event dataset identifies 68 episodes under SSP1-2.6 and 63 under SSP5-8.5. Event count alone is therefore not a sufficient indicator of worsening drought: under SSP5-8.5, episodes are longer on average (12.8 months versus 10.9 months) and more severe (mean cumulative severity of 25.3 versus 20.4) (Table 3 and Figure 5). Under SSP1-2.6, the longest continuous episode extends from June 2062 to April 2067, or 59 months. Under SSP5-8.5, several sequences reach 35 months, including 2047–2050, 2051–2054, 2056–2059, and 2096–2099. The succession of long dry sequences under SSP5-8.5 confirms the tendency toward chronic drought indicated by monthly frequencies. Table 3. Summary of hydrological drought events (SSI-3 ≤ −1) over 2021–2100. Scenario Number Mean duration (months) Maximum duration Mean severity Longest event SSP1-2.6 68 10.9 59 20.4 2062-06-01 to 2067-04-01 SSP5-8.5 63 12.8 35 25.3 2047-06-01 to 2050-04-01 Figure 5. Distribution of hydrological drought-event duration. 3.4 Performance of SSI-3 Forecasting Models At a one-month lead time, boosting methods are particularly effective. LightGBM achieves the highest NSE under both scenarios: 0.867 under SSP1-2.6 (KGE = 0.870; RMSE = 0.313) and 0.718 under SSP5-8.5 (KGE = 0.693; RMSE = 0.308). Under SSP1-2.6, the other methods are also robust at one month: ExtraTrees reaches an NSE of 0.863, XGBoost 0.858, Random Forest 0.854, and GRU 0.850. The closeness of these scores indicates that recent system memory and lagged variables contain a strong predictive signal (Table 4 and Figure 6). Table 4. Best model by scenario and forecast horizon according to test-set NSE. Scenario Lead time Model NSE KGE RMSE MAE F1 SSP1-2.6 1 LightGBM 0.867 0.870 0.313 0.183 0.960 SSP1-2.6 3 LSTM 0.713 0.671 0.459 0.314 0.951 SSP1-2.6 6 GRU 0.539 0.591 0.582 0.442 0.887 SSP5-8.5 1 LightGBM 0.718 0.693 0.308 0.179 0.978 SSP5-8.5 3 GRU 0.175 0.135 0.528 0.283 0.951 SSP5-8.5 6 LSTM 0.077 0.066 0.558 0.370 0.951 At three months under SSP1-2.6, LSTM becomes the best model with NSE = 0.713 and F1 = 0.951, followed by GRU (NSE = 0.692) and MLP/ANN (0.676). At six months, GRU retains the highest NSE (0.539), slightly ahead of LSTM (0.527) and MLP/ANN (0.501). This evolution is consistent with the value of recurrent architectures when the forecast horizon requires representation of longer temporal dependencies. Under SSP5-8.5, forecasting difficulty increases sharply with lead time. At three months, the best NSE is only 0.175 for GRU; at six months, LSTM achieves 0.077. Several models even produce negative NSE values. Nevertheless, F1 scores remain close to 0.95 for the best recurrent networks. This contrast means that the models correctly recognize the drought state, which becomes very frequent, but have greater difficulty reproducing variations in SSI-3 intensity. In a context where more than 90% of months are dry by the end of the century, a high F1 score must therefore be interpreted jointly with NSE, KGE, and continuous error metrics. Figure 6. NSE of the eight models by scenario and forecast horizon. 3.5 Variables Controlling Forecast Skill At a one-month lead time, persistence in hydrological indices clearly dominates. Under SSP1-2.6, SSI3_lag0 accounts for 35.0% of LightGBM importance, SSI1_lag0 for 28.5%, and SSI1_lag1 for 7.7%. Under SSP5-8.5, SSI1_lag0 reaches 45.9%, SSI3_lag0 20.5%, followed by Discharge_lag2 at 6.7% and SSI1_lag1 at 5.8% (Figure 7). Short-term forecasting therefore depends primarily on the immediately preceding hydrological state. As the forecast horizon increases, the role of climate forcing and streamflow memory becomes stronger. Under SSP1-2.6 at three months, Discharge_lag3 is the leading variable (25.2%), followed by SSI1_lag0 (12.7%) and Precipitation_lag6 (8.6%). At six months, Precipitation_lag2 becomes dominant (20.6%), ahead of Discharge_lag1 (10.4%). Under SSP5-8.5, precipitation lagged by two months is also the leading variable at six months (16.0%), followed by current streamflow (10.6%). These results highlight a transition from hydrological persistence at very short lead times to multi-timescale hydroclimatic control at seasonal horizons. Figure 7. Main predictor importance values for the LightGBM model at a one-month lead time.

07Discussion

The results reveal a marked divergence between meteorological, climatic, and hydrological drought responses in the Gambia River basin at Kédougou. While SPI-3 remains comparatively moderate across several future periods, SPEI-3 and especially SSI-3 show substantially stronger negative anomalies. This confirms that precipitation alone is insufficient to characterize future drought conditions, particularly under global warming, because drought results from interactions among precipitation deficits, atmospheric evaporative demand, catchment storage, and runoff processes (Mishra & Singh, 2010; Vicente-Serrano et al., 2010). The stronger deterioration of SPEI compared with SPI is therefore physically consistent with increasing temperature and potential evapotranspiration, while the pronounced SSI response demonstrates the propagation and amplification of climatic anomalies through the hydrological system (Vicente-Serrano et al., 2012; Wang et al., 2022). This amplification becomes particularly evident under SSP5-8.5. Between 2021–2040 and 2081–2100, mean temperature increases by 4.32 °C and PET by 12.4%, whereas mean streamflow decreases by approximately 74.6%, despite the absence of a comparable monotonic decline in precipitation. Such nonlinear rainfall–runoff behavior is consistent with the fundamental principles of catchment hydrology, whereby streamflow integrates precipitation, evapotranspiration, antecedent moisture conditions, storage, and basin characteristics (Beven, 2012). Comparable sensitivities have been reported for the Gambia basin, where rainfall–runoff simulations have demonstrated the importance of climatic forcing in controlling discharge variability (Faty et al., 2018; Faty et al., 2023). The projected contraction of flow also agrees with previous assessments indicating substantial climate-induced changes in the Senegal and Gambia river basins (Bodian et al., 2018; Rameshwaran et al., 2021). The contrast between SSP1-2.6 and SSP5-8.5 further demonstrates the importance of emission pathways. Although SSP1-2.6 does not eliminate hydrological drought, it avoids the almost continuous deterioration projected under SSP5-8.5. This scenario dependence is consistent with the conceptual foundations of the Shared Socioeconomic Pathways and ScenarioMIP framework (Riahi et al., 2017; O'Neill et al., 2016). It is also consistent with recent basin-scale assessments showing that future hydroclimatic conditions in the Gambia basin are highly sensitive to climate forcing (Séne et al., 2024; Kah et al., 2025). The increasing frequency and persistence of extreme drought under SSP5-8.5 consequently suggest a transition from episodic drought toward quasi-chronic hydrological stress. The forecasting results provide an additional methodological contribution. LightGBM achieves the highest performance at a one-month lead time, with NSE values of 0.867 under SSP1-2.6 and 0.718 under SSP5-8.5. This performance is consistent with the capacity of tree-based ensemble methods to capture nonlinear interactions among hydroclimatic predictors, as demonstrated more broadly for Random Forest approaches (Breiman, 2001). At longer horizons, however, LSTM and GRU become comparatively more effective, particularly under SSP1-2.6, confirming the relevance of recurrent architectures for learning temporal dependencies and hydrological memory (Zhao et al., 2024). Similar advantages of combining nonlinear machine-learning approaches with temporal information have been reported for long-term streamflow prediction (Abbasi et al., 2022). Nevertheless, predictive skill deteriorates markedly at three- and six-month horizons under SSP5-8.5. This decline may reflect increasing hydroclimatic non-stationarity and the difficulty of reproducing SSI intensity when future conditions increasingly depart from the historical 1985–2014 reference distribution. Although F1 scores remain high, the low NSE values indicate that successful drought-state classification does not necessarily imply accurate prediction of drought magnitude. Model performance must therefore be evaluated using complementary metrics rather than a single criterion (Moriasi et al., 2007; Gupta et al., 2009). Finally, several uncertainties remain. Future SSI values depend on simulated streamflow, meaning that hydrological-model uncertainty propagates into drought characterization. Climate projections also contain uncertainties associated with model structure, emission scenarios, and regional climate representation (Abdulai & Chung, 2019; Almeida et al., 2024). Moreover, changes in land use can modify rainfall–runoff relationships and potentially reinforce or attenuate climate-driven hydrological responses (Ndiaye et al., 2024; Ba et al., 2024). These limitations support the future use of multi-GCM, multi-hydrological-model and ensemble forecasting frameworks. Nevertheless, the convergence of drought indices, streamflow projections, and artificial-intelligence models provides strong evidence that the upper Gambia River basin may face substantially greater hydrological drought pressure during the twenty-first century, particularly under high greenhouse-gas emissions. The results suggest that future management of the upper Gambia River basin will need to address both declining mean streamflow and longer drought sequences. Under SSP5-8.5, the very high frequency of dry months by the end of the century implies potential constraints on water supply, agricultural uses, aquatic ecosystems, and river dilution capacity. Adaptation strategies should therefore strengthen hydroclimatic monitoring, conservation of recharge areas and riparian zones, water-use efficiency, seasonally appropriate storage, and mechanisms for transboundary information sharing. From an early-warning perspective, the performance obtained at one month supports the development of an operational system combining LightGBM with a recurrent neural network. At three to six months, a probabilistic or ensemble approach would be preferable, particularly under SSP5-8.5. The strong importance of lagged SSI and streamflow indicates that an operational system must maintain continuous hydrometric observations; precipitation and PET become increasingly important for extending the forecast horizon.

08Conclusion

This study provides an integrated assessment of future drought dynamics and predictability in the Gambia River basin at Kédougou by combining meteorological, climatic, and hydrological drought indices with machine-learning and deep-learning approaches. The joint analysis of SPI, SPEI, and SSI demonstrates that future drought cannot be adequately assessed from precipitation variability alone. The stronger deterioration of SPEI and, particularly, SSI highlights the critical influence of increasing evaporative demand and the amplification of climatic anomalies through the hydrological system. The results reveal contrasting trajectories under SSP1-2.6 and SSP5-8.5. Although substantial hydrological drought persists under the low-emission pathway, SSP5-8.5 produces a pronounced and progressive deterioration of river-flow conditions. By 2081–2100, mean temperature reaches 33.99 °C, annual PET increases to 1,770.3 mm, and mean streamflow declines to 12.45 m³/s. Concurrently, mean SSI-3 falls to −1.94, while 91.7% of months are classified as drought and 84.2% as extreme drought. These changes indicate a potential transition from episodic drought toward quasi-chronic hydrological drought by the end of the century. The forecasting experiment further demonstrates that model performance depends strongly on forecast horizon and climate scenario. LightGBM provides the best one-month forecasts, reaching NSE values of 0.867 and 0.718 under SSP1-2.6 and SSP5-8.5, respectively. At longer lead times, LSTM and GRU become more competitive, particularly under SSP1-2.6. However, the sharp decline in continuous predictive performance at three and six months under SSP5-8.5 highlights the challenge posed by increasing hydroclimatic non-stationarity. Overall, these findings demonstrate the value of combining standardized drought indicators with multimodel artificial-intelligence forecasting for drought early warning. They also emphasize the need to strengthen hydrometric monitoring, incorporate multi-model climate and hydrological uncertainties, and develop adaptive water-management strategies. Limiting greenhouse-gas emissions and improving anticipatory water governance appear complementary for reducing the risk of severe and persistent hydrological drought in the Gambia River basin.

03References

35 bibliographic references

  1. Abdulai, P. J., & Chung, E. (2019). Uncertainty assessment in drought severities using multiple GCMs and the reliability ensemble averaging method. Sustainability.

  2. Ahmadalipour, A., Moradkhani, H., & Demirel, M. (2017). A comparative assessment of projected meteorological and hydrological droughts: Elucidating the role of temperature. Journal of Hydrology, 553, 785–797.

  3. Beven, K. J. (2012). Rainfall-runoff modelling: The primer (2nd ed.). Wiley-Blackwell.

  4. Bodian, A., Dezetter, A., Diop, L., Déme, A., Djaman, K., & Diop, A. (2018). Future climate change impacts on streamflows of two main West Africa river basins: Senegal and Gambia. Hydrology, 5, 21.

  5. Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.

  6. Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). Association for Computing Machinery. https://doi.org/10.1145/2939672.2939785

  7. Cho, K., van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning phrase representations using RNN encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) (pp. 1724–1734). Association for Computational Linguistics. https://doi.org/10.3115/v1/D14-1179

  8. Dione, O. (1996). Évolution climatique récente et dynamique fluviale dans les hauts bassins des fleuves Sénégal et Gambie [Doctoral dissertation, Université Jean Moulin Lyon 3].

  9. Diop, S. B., Ekolu, J., Tramblay, Y., Dieppois, B., Grimaldi, S., Bodian, A., et al. (2025). Climate change impacts on floods in West Africa: New insight from two large-scale hydrological models. Natural Hazards and Earth System Sciences, 25(9), 3161–3184.

  10. Faty, B., Ali, A., Dacosta, H., Bodian, A., Diop, S., & Descroix, L. (2018). Assessment of satellite rainfall products for stream flow simulation in Gambia watershed. African Journal of Environmental Science and Technology, 12, 501–513.

  11. Faty, B., Sterk, G., Ali, A., Sy, S., Dacosta, H., Diop, S., & Descroix, L. (2023). Satellite-based rainfall estimates to simulate daily streamflow using a hydrological model over Gambia watershed. Water Science, 37, 151–168.

  12. Faye, C., & Mendy, A. (2018). Climate variability and hydrological impacts in West Africa: The case of the Gambia River basin (Senegal). Environmental and Water Sciences, Public Health.

  13. Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G., Rowland, J., Harrison, L., Hoell, A., & Michaelsen, J. (2015). The climate hazards infrared precipitation with stations—A new environmental record for monitoring extremes. Scientific Data, 2, Article 150066. https://doi.org/10.1038/sdata.2015.66

  14. Gupta, H. V., Kling, H., Yilmaz, K. K., & Martinez, G. F. (2009). Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. Journal of Hydrology, 377(1-2), 80–91.

  15. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735

  16. Kah, M., Faye, C., Mbaye, M. L., Yalo, N., & Lischeid, G. (2025). Hydroclimatic trends and land use changes in the continental part of the Gambia River Basin: Implications for water resources. Water, 17(14), 2075.

  17. Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., & Liu, T.-Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 30, 3146–3154.

  18. Liu, T., Si, Z., Liu, Y., et al. (2025). Runoff and drought responses to land use change and CMIP6 climate projections. Water.

  19. McKee, T. B., Doesken, N. J., & Kleist, J. (1993). The relationship of drought frequency and duration to time scales. In Proceedings of the 8th Conference on Applied Climatology (pp. 179–184). American Meteorological Society.

  20. Mishra, A. K., & Singh, V. P. (2010). A review of drought concepts. Journal of Hydrology, 391, 202–216.

  21. Moriasi, D. N., Arnold, J. G., Van Liew, M. W., Bingner, R. L., Harmel, R. D., & Veith, T. L. (2007). Model evaluation guidelines for systematic quantification of accuracy in watershed simulations. Transactions of the ASABE, 50(3), 885–900.

  22. Muñoz-Sabater, J., Dutra, E., Agustí-Panareda, A., Albergel, C., Arduini, G., Balsamo, G., et al. (2021). ERA5-Land: A state-of-the-art global reanalysis dataset for land applications. Earth System Science Data, 13(9), 4349–4383. https://doi.org/10.5194/essd-13-4349-2021

  23. O'Neill, B. C., Tebaldi, C., van Vuuren, D. P., Eyring, V., Friedlingstein, P., Hurtt, G., et al. (2016). The Scenario Model Intercomparison Project (ScenarioMIP) for CMIP6. Geoscientific Model Development, 9(9), 3461–3482.

  24. Oudin, L., Hervieu, F., Michel, C., Perrin, C., Andréassian, V., Anctil, F., & Loumagne, C. (2005). Which potential evapotranspiration input for a lumped rainfall–runoff model? Part 2—Towards a simple and efficient potential evapotranspiration model for rainfall–runoff modelling. Journal of Hydrology, 303(1–4), 290–306. https://doi.org/10.1016/j.jhydrol.2004.08.026

  25. Rameshwaran, P., Bell, V. A., Davies, H. N., & Kay, A. L. (2021). How might climate change affect river flows across West Africa? Climatic Change, 169, 21.

  26. Riahi, K., van Vuuren, D. P., Nakicenovic, N., Fricko, O., Krey, V., Gidden, M., et al. (2017). The Shared Socioeconomic Pathways and their energy, land use, and greenhouse gas emissions implications: An overview. Global Environmental Change, 42, 153–168.

  27. Séne, S. M. K., Faye, C., & Pande, C. B. (2024). Assessment of current and future trends in water resources in the Gambia River Basin in a context of climate change. Environmental Sciences Europe, 36, 32.

  28. Shukla, S., & Wood, A. W. (2008). Use of a standardized runoff index for characterizing hydrologic drought. Geophysical Research Letters, 35, L02405.

  29. Song, Y., Shahid, S., & Chung, E. (2021). Differences in multi-model ensembles of CMIP5 and CMIP6 projections for future droughts in South Korea. International Journal of Climatology, 42, 2688–2716.

  30. Sow, A. A. (2007). L'hydrologie du Sud-Est du Sénégal et de ses confins guinéo-maliens : Les bassins de la Gambie et de la Falémé [Doctoral dissertation, Université Cheikh Anta Diop de Dakar].

  31. Traore, V., Sambou, S., Cisse, M., et al. (2014). Trends and shifts in time series of rainfall and runoff in the Gambia River Watershed. International Journal of Environmental Protection and Policy, 2, 138.

  32. Vicente-Serrano, S. M., Beguería, S., & López-Moreno, J. I. (2010). A multiscalar drought index sensitive to global warming: The Standardized Precipitation Evapotranspiration Index. Journal of Climate, 23(7), 1696–1718.

  33. Vicente-Serrano, S. M., López-Moreno, J. I., Beguería, S., Lorenzo-Lacruz, J., Azorin-Molina, C., & Morán-Tejeda, E. (2012). Accurate computation of a streamflow drought index. Journal of Hydrologic Engineering, 17(2), 318–332. https://doi.org/10.1061/(ASCE)HE.1943-5584.0000433

  34. Wang, Y., et al. (2022). Drought propagation under global warming: Characteristics, approaches, processes, and controlling factors. Science of the Total Environment, 838, 156021.

  35. Zhao, X., Wang, H., Bai, M., Xu, Y., Dong, S., Rao, H., & Ming, W. (2024). A comprehensive review of methods for hydrological forecasting based on deep learning. Water, 16(10), 1407.

20Views
88Downloads

© 2026 The Authors. Published by EcoClean Scientific Publishing on behalf of JEERESD. Distributed under the Creative Commons Attribution 4.0 International License.