To assess the impact of climate change on surface water availability in the Diaguiri River Basin, we compiled a comprehensive dataset of observed hydroclimatic variables, including daily rainfall, temperature, and river discharge records from multiple stations distributed across the basin ( Table 1 ). The main hydrometric station is located at the Diaguiri road bridge (Lat: 12.63˚N, Long: −12.08˚W), where daily river discharge was recorded from June 15, 1974 to December 31, 2005, with a mean observed flow of 6.30 m³/s. Rainfall observations were obtained from six meteorological stations within and around the basin, covering various time periods from 1950 to 2016, with mean annual rainfall ranging from 1050 mm at Nepen Diakha to 1233 mm at Kédougou. Temperature data were available from the Kédougou station for the period 1972-2012, with a mean annual temperature of 29˚C. Potential evapotranspiration (PET) was estimated using the method proposed by Oudin (2005) , resulting in an average PET of 1903 mm/year across the basin. The spatially representative average rainfall across the Diaguiri Basin was derived using Kriging interpolation techniques implemented in the Hydraccess software ( Vauchel , 2004 ). The use of a single temperature station to estimate Potential Evapotranspiration (PET) in the Diaguiri River Basin relies on the assumption that temperature variations at this station adequately represent those across the entire basin. However, this does not account for localized microclimates or spatial temperature variability, which could introduce uncertainty in PET estimation.

The GR4J (Génie Rural à 4 paramètres Journalier) model is a daily lumped rainfall-runoff model developed by Perrin et al. (2003) at the French research institute CEMAGREF (now INRAE). The model adopts a parsimonious approach, relying on just four parameters to characterize the hydrological behavior of a watershed. Its conceptual simplicity and ability to simulate streamflow effectively in catchments with limited hydro-meteorological data make it especially suitable for application in data-scarce watersheds like the Diaguiri River Basin in West Africa. The model employs a parsimonious structure, characterized by only four parameters that capture the essential hydrological behavior of a watershed. These parameters include the production store capacity (X1), which defines the maximum water volume that the soil reservoir can hold; the groundwater exchange coefficient (X2), which accounts for water exchanges with deeper groundwater or adjacent basins; the routing store capacity (X3), which controls the retention of flow within the routing reservoir; and the unit hydrograph time base (X4), which governs the timing and shape of the hydrograph ( Figure 2 ). Operating at a daily timestep, the model requires daily inputs of precipitation (P) and potential evapotranspiration (PET). Rather than spatially discretizing the catchment, GR4J treats it as a lumped, homogeneous unit. Key hydrological processes—including evapotranspiration, infiltration, percolation, and runoff routing—are conceptualized through storage components and transfer functions. The model’s architecture features a soil moisture production store that simulates the partitioning of precipitation into net evapotranspiration capacity (En) and effective rainfall (Pn). This effective rainfall is subsequently split into two flow components: one routed through a unit hydrograph producing rapid runoff (UH1), and the other through a secondary hydrograph responsible for delayed flow (UH2). These flow components are combined and routed through a routing store (RS), which simulates groundwater discharge and attenuates flow peaks ( Figure 2 ).

The calibration and validation of the GR4J hydrological model were carried out using observed daily rainfall (P), Potential Evapotranspiration (PET), and streamflow data (Q) collected from the Diaguiri River Basin. The calibration phase was conducted over the period 1981-1992, while the model validation was performed for the period 1998-2004. To minimize the influence of initial conditions on model outputs, warm-up periods were introduced in 1980 for the calibration phase and 1997 for the validation phase. The exclusion of discharge data from 1993 to 1997 is justified by gaps in the time series or data quality issues that could have skewed the results. By excluding these years, only complete and reliable data are used for calibration and validation, thereby strengthening the robustness of the cross-validation approach and the reliability of future hydrological projections. The model optimization was performed using the Shuffled Complex Evolution-University of Arizona (SCE-UA) algorithm ( Duan et al., 1992 ), which is

widely recognized for its robustness in exploring complex parameter spaces and avoiding local optima in hydrological model calibration ( Vrugt et al., 2003 ; Thiaw, 2020 ; Thiaw et al., 2021 ). The GR4J model, with its parsimonious structure, requires the careful estimation of its four parameters to adequately capture the hydrological response of the Diaguiri watershed under both historical and future climatic conditions ( Perrin et al., 2003 ; Pushpalatha et al., 2012 ; Thiaw, 2020 ). The use of observed data and a rigorous optimization technique ensures the reliability of the calibrated model for subsequent climate change impact assessments on surface water availability in this West African basin.

In this study, the evaluation of the GR4J model’s performance in the Diaguiri River Basin was conducted using a multi-objective function that integrates four complementary performance metrics: Nash-Sutcliffe Efficiency (NSE), log-transformed Nash-Sutcliffe Efficiency (log-NSE), Kling-Gupta Efficiency (KGE), and the coefficient of determination (R²). The combined use of these metrics allows for a more comprehensive and balanced assessment of the model’s ability to reproduce both high-flow and low-flow conditions under varying hydrological regimes ( Nash & Sutcliffe, 1970 ; Gupta et al., 2009 ; Pushpalatha et al., 2012 ; Kling et al., 2012 ). The inclusion of both NSE and log-NSE is especially important to capture the model’s sensitivity to extreme flows as well as baseflows, particularly in a watershed characterized by strong seasonality and alternating periods of floods and droughts. NSE evaluates the overall agreement between observed and simulated streamflows, with values approaching 1 indicating better model performance ( Nash & Sutcliffe, 1970 ; Moriasi et al., 2007 ). The log-NSE, by applying a logarithmic transformation to both observed and simulated flows, enhances the evaluation of low-flow conditions that are critical for water availability during dry seasons ( Pushpalatha et al., 2012 ). The KGE metric, proposed by Gupta et al. (2009) , integrates correlation (r), bias (β), and variability (α), providing a global diagnostic of model behavior across different flow regimes. Finally, R² measures the strength of linear correlation between observed and simulated discharges, offering a simple indicator of the model’s overall fit ( Legates & McCabe, 1999 ). The corresponding mathematical formulations of these performance metrics are as follows:

$NSE=1-\frac{\left[\Sigma {\left(Q_{obs,i}-Q_{sim,i}\right)}^2\right]}{\left[\Sigma {\left(Q_{obs,i}-{\bar{Q}}_{obs}\right)}^2\right]}$ (1)

$R^2=\frac{{\left[\Sigma \left(\left(Q_{obs,i}-{\bar{Q}}_{obs}\right)\left(Q_{sim,i}-{\bar{Q}}_{sim}\right)\right)\right]}^2}{\left[\Sigma {\left(Q_{obs,i}-{\bar{Q}}_{obs}\right)}^2\ast \Sigma {\left(Q_{sim,i}-{\bar{Q}}_{sim}\right)}^2\right]}$ (2)

$\log -NSE=1-\frac{\left[\Sigma {\left(\log \left(Q_{obs,i}\right)-\log \left(Q_{sim,i}\right)\right)}^2\right]}{\left[\Sigma {\left(\log \left(Q_{obs,i}\right)-\log \left({\bar{Q}}_{obs}\right)\right)}^2\right]}$(3)

$KGE=1-\sqrt{\left[{\left(r-1\right)}^2+{\left(\alpha -1\right)}^2+{\left(\beta -1\right)}^2\right]}$ (4)

where Q_obs,i and Q_sim,i are the observed and simulated streamflows at time step i, respectively; ${\bar{Q}}_{obs}$ and ${\bar{Q}}_{sim}$ are their corresponding mean values; r is the linear correlation coefficient; α is the variability ratio (i.e., the ratio of standard deviations σ_sim/σ_obs); and β is the bias ratio μ_sim/μ_obs.

The adoption of multiple performance indicators allows the model evaluation to capture the full range of hydrological variability observed in the Diaguiri watershed. Moreover, the integration of the GR4J model with the SCE-UA optimization algorithm ( Duan et al., 1992 ) within airGR ( Coron et al., 2017 ) provides a robust and adaptable framework, particularly well-suited for simulating climate-change scenarios using bias-corrected outputs from Regional Climate Models (RCMs).

Future climate scenarios were developed using bias-corrected outputs from a suite of Regional Climate Models (RCMs) produced under the CORDEX-Africa initiative (Coordinated Regional Downscaling Experiment for Africa), which provides dynamically downscaled simulations at approximately 0.44˚ resolution suitable for basin-scale climate impact assessments in Africa ( Giorgi et al., 2009 ; Nikulin et al., 2012 ). These RCMs, driven by different Global Climate Models (GCMs) under the RCP4.5 and RCP8.5 scenarios, were selected for their demonstrated skill in representing the seasonal and interannual variability of the West African monsoon ( Nikulin et al., 2012 ; Sylla et al., 2013 ; Dosio et al., 2015 ; Diallo et al., 2016 ; Thiaw et al., 2021 ). The table below summarizes the climate models used in this study.

To enhance the physical realism of the climate projections, the raw RCM outputs were bias-corrected using the CDF-t method (Cumulative Distribution Function transform), which adjusts both the mean and higher-order moments of the modeled distribution to match those of observed data while preserving climate change signals ( Michelangeli et al., 2009 ; Vrac et al., 2012 ). This correction technique has been widely validated and is particularly suited for applications in hydrological impact modeling under climate change conditions ( Teutschbein & Seibert, 2012 ; Thiaw, 2020 ; Thiaw et al., 2021 ). The bias correction utilized observed daily data of precipitation (1950-2016) and temperature (1972-2012) from the Diaguiri basin for the historical reference periods. Each climate variable was corrected independently, and the procedure was applied at the scale of each grid cell intersecting the catchment. To aggregate the bias-corrected RCM precipitation grids to basin-scale inputs, the Kriging method was employed using Hydraccess software ( Vauchel , 2004 ). The resulting bias-adjusted time series were then used as inputs to simulate streamflow responses over the future period 2021-2100, which was subdivided into three-time horizons: near-term (2021-2040), mid-century (2041-2070), and end-of-century (2071-2100).

The evaluation of the GR4J model’s performance during the cross-validation phase highlights its robustness in simulating streamflow within the semi-humid Diaguiri River Basin. The calibration, conducted over the period 1981-1992, and the subsequent validation, carried out for 1998-2004, enabled an assessment of the model’s transferability across distinct climatic conditions. During calibration, the Nash-Sutcliffe Efficiency (NSE) reached 0.66 and improved to 0.71 during validation, indicating good agreement between observed and simulated flows, particularly during the rainy season when runoff generation is most pronounced ( Figure 3 and Figure 4 ). The log-transformed NSE (log-NSE), which emphasizes low-flow dynamics, increased markedly from 0.54 to 0.77, suggesting that the model effectively reproduces both high-flow events and baseflows ( Figure 3 and Figure 4 )—an essential aspect in semi-humid basins where streamflow is governed by the alternation between wet and dry seasons ( Thiaw, 2020 ; Thiaw et al., 2021 ). The Kling-Gupta Efficiency (KGE), which integrates correlation, bias, and variability, also showed improvement from 0.71 to 0.83, confirming the model’s balanced performance across hydrological components. Similarly, the coefficient of determination (R²) followed the same trend as the NSE (0.66 and 0.71), further attesting to the reliability of the simulations ( Figure 3 and Figure 4 ).

These findings confirm the GR4J model’s reliability and suitability for projecting future streamflow under climate change scenarios, thereby reinforcing confidence in the anticipated hydrological shifts within the Diaguiri River Basin. In light of the model’s stable and consistent performance, particularly during the validation phase (1998-2004), the corresponding parameter set—X1: 207.40 mm, X2: −30.66 mm, X3: 333.28 mm, and X4: 2.09 days—was retained for future simulations.

The CDF-t bias correction technique was applied separately to the temperature outputs of each climate model, using observed daily data from 1972 to 2012 as a reference period. This approach allowed for the adjustment of the modeled temperature distributions so that they better reflected the statistical characteristics of observed series, notably their mean values and interannual variability. Under the RCP4.5 emissions scenario, bias correction led to noticeable improvements across most models. For instance, GFDL-ESM2M reproduced the observed standard deviation relatively well (2.09˚C versus 2.23˚C observed), with a correlation of 0.73 and a root mean square deviation (RMSD) of 2.12˚C. HadGEM2-ES and IPSL-CM5A-LR performed even better, with correlation coefficients of 0.89 and 0.91, respectively, and RMSD values below 2˚C. Although MIROC5 slightly exaggerated the variability, it retained a low RMSD of 1.65˚C. NorESM1-M yielded the most consistent results, combining the lowest RMSD (1.38˚C) with a high correlation of 0.92. Under the RCP8.5 scenario, model behavior remained broadly similar. GFDL-ESM2M and HadGEM2-ES showed higher RMSDs (2.29˚C and 2.33˚C), while IPSL-CM5A-LR and MIROC5 maintained strong correlation values (0.89) and RMSDs of 2.35˚C and 1.54˚C, respectively. Once again, NorESM1-M provided the most stable results, with a correlation of 0.91 and an RMSD of 1.68˚C ( Figure 5 ).

Overall, the CDF-t method significantly improved the alignment between modeled and observed temperature data. Among the ensemble, NorESM1-M, MIROC5, and IPSL-CM5A-LR emerged as the most reliable performers under both RCP4.5 and RCP8.5 scenarios. These improvements in temperature accuracy reinforce the credibility of hydrological projections for the Diaguiri Basin, thereby contributing to more robust planning for climate adaptation and sustainable water management in this key sub-basin of the Gambia River.

The CDF-t method was also applied to precipitation outputs from multiple climate models, using observed rainfall data from 1950 to 2016 in the Diaguiri River Basin as a reference. This bias correction markedly improved the representation of the statistical properties of precipitation, especially its variability and distribution, which are critical for accurately modeling runoff generation and hydrological extremes in climate-sensitive catchments. Under the RCP4.5

scenario, MIROC5 and NorESM1-M showed the best agreement with observed precipitation patterns. Their standard deviations closely matched the observed value of 3.76 mm, with high correlation coefficients of 0.86 and 0.83, respectively, and the lowest root mean square deviations (RMSDs) of 2.22 mm and 2.20 mm. IPSL-CM5A-LR and HadGEM2-ES also performed well, though HadGEM2-ES slightly underestimated interannual variability. GFDL-ESM2M tended to overestimate variability but maintained a reasonable correlation of 0.80, with an RMSD of 2.51 mm. Under the RCP8.5 scenario, overall model performance remained consistent. HadGEM2-ES and NorESM1-M continued to deliver strong results, with RMSDs near 2.15 mm and correlations above 0.83. MIROC5 slightly overestimated variability (4.32 mm) but retained a strong correlation of 0.82. GFDL-ESM2M again had the highest RMSD (2.63 mm), reflecting persistent overestimation, while IPSL-CM5A-LR showed balanced moderate performance ( Figure 6 ). These findings highlight the essential role of bias correction in refining climate model outputs for hydrological simulations. Given precipitation’s key role as the main driver of hydrological processes in the Diaguiri River Basin, these corrected datasets substantially improve the reliability of simulated runoff, streamflow, and hydrological extremes under future climate scenarios.

The projected temperature trends derived from the multi-model ensemble reveal a consistent and progressive warming of the Diaguiri River Basin throughout the 21st century under both RCP4.5 and RCP8.5 scenarios. Under the intermediate stabilization pathway RCP4.5, the mean temperature anomaly relative to the historical baseline increases from +1.29˚C during the near-term period (2021-2040) to +2.01˚C in the mid-century (2041-2070), reaching +2.41˚C by the end of the century (2071-2100) ( Figure 7 ). This trajectory reflects a gradual warming trend, aligning with global efforts to limit radiative forcing and greenhouse gas emissions after mid-century. In contrast, under the high-emission RCP8.5 scenario—which assumes continued growth in emissions—the projected warming is more pronounced and accelerates over time. Temperature anomalies begin at +1.45˚C in the near future (2021-2040), rise sharply to +2.66˚C by mid-century (2041-2070), and culminate at a significant +4.37 °C by the late 21st century (2071-2100) ( Figure 8 ). This steeper trajectory under RCP8.5 indicates a strong sensitivity of the regional climate system to greenhouse gas forcing, with the latter period exhibiting nearly double the warming compared to RCP4.5.

These results highlight the growing divergence between scenarios after mid-century and underscore the critical influence of future emission pathways on regional climate outcomes. The projected warming—especially under RCP8.5—has profound implications for the hydrological regime of the Diaguiri Basin, including increased evapotranspiration, altered runoff patterns, and heightened stress on water resources.

The projected evolution of precipitation over the Diaguiri River Basin throughout the 21st century, based on bias-corrected outputs from a suite of five CMIP5 global climate models (GFDL-ESM2M, HadGEM2-ES, IPSL-CM5A-LR, MIROC5, and NorESM1-M), reveals marked temporal heterogeneity and pronounced scenario-dependent divergence. Under the RCP4.5 stabilization pathway, ensemble projections indicate a modest but progressively increasing hydroclimatic signal. Mean annual precipitation anomalies relative to the historical baseline (1956-2016) are projected to rise by +6.72% during the near term (2021-2040), with a slight attenuation to +5.79% in the mid-century (2041-2070), followed by a more substantial increase of +12.16% toward the end of the century (2071-2100). This gradual upward trajectory suggests a potential intensification of moisture availability under moderate radiative forcing. However, the inter-model spread remains substantial—ranging from −11.96% (IPSL-CM5A) to +18.48% (HadGEM2-ES) in the near term, and reaching up to +39.89% (MIROC5) by century’s end ( Figure 9 ). Conversely, the RCP8.5 high-emissions scenario exhibits a more erratic and nonlinear precipitation response, characterized by amplified inter-model variance and weakened convergence. The ensemble mean anomaly peaks early in the century at +13.68% (2021-2040), reflecting a greenhouse gas-driven intensification of the regional hydrological cycle. This signal diminishes to +10.19% by mid-century and further weakens to +5.32% by 2071-2100, despite highly divergent individual model outputs. For instance, while GFDL-ESM2M projects an extreme increase of +62.7% by century’s end, HadGEM2-ES and IPSL-CM5A-LR simulate pronounced drying trends of −31.7% and −28.02%, respectively ( Figure 10 ). This divergence points to deep model disagreement under strong radiative forcing, likely driven by disparate representations of deep convection, land-atmosphere feedbacks, and large-scale circulation shifts affecting monsoonal dynamics.

Overall, these findings underscore the profound uncertainty inherent in long-term precipitation projections for the Diaguiri River Basin, particularly under RCP8.5. While RCP4.5 supports a more coherent signal of modest wetting, the high-emissions trajectory entails a shift toward hydroclimatic non-stationarity, with potential for both substantial intensification and severe rainfall deficits depending on the model and time horizon. Given the basin’s strong precipitation-runoff coupling, such variability presents critical implications for future hydrological responses, water availability, and flood risk.

Under the RCP4.5 stabilization pathway, long-term projections indicate a substantial intensification of annual runoff volumes in the Diaguiri River Basin over the 2021-2100 period. The multi-model ensemble mean (MMM) streamflow is projected to increase by +24% relative to the historical baseline, a response that closely reflects the basin’s precipitation-driven hydrological regime ( Table 3 , Figure 11 ). This enhanced runoff signal aligns with the ensemble mean precipitation anomaly of +8.37%, based on bias-corrected outputs from five CMIP5 global climate models. However, considerable inter-model variability underscores the system’s nonlinear rainfall-runoff response. For instance, MIROC5, which simulates the highest precipitation anomaly (+23.45%, equivalent to +273 mm), yields the most pronounced streamflow amplification (+50.7%), while GFDL-ESM2M follows closely with a +15.6% rainfall increase driving a +45.6% discharge gain. NorESM1-M, with a moderate precipitation rise of +10.0%, produces a corresponding +33% increase in mean annual flow. In contrast, HadGEM2-ES projects a near-stationary precipitation signal (−1.35%) and a marginal discharge increase (+3.9%), while IPSL-CM5A-LR, exhibiting a drying trend (−5.88%), simulates a negative runoff response (−12.9%) ( Figure 11 ). These divergent projections emphasize the Diaguiri Basin’s strong hydroclimatic sensitivity and the critical role of antecedent moisture conditions and model-specific rainfall dynamics in modulating future streamflow. The amplification of discharge variability, evident through increases in ensemble spread and higher-order flow statistics, further reflects the basin’s potential transition toward hydrological non-stationarity under evolving climate conditions.

While the RCP4.5 scenario outlines a moderately wetter future with relatively consistent precipitation-runoff responses, the RCP8.5 pathway introduces a more complex and heterogeneous hydrological signal, driven by amplified radiative forcing and divergent precipitation trajectories across models. Indeed, by 2100,

the multi-model mean (MMM) annual streamflow increases by +34.3%, rising from a historical baseline of 6.20 m³/s to 8.30 m³/s. This hydrological amplification corresponds to a +9.3% increase in multi-model ensemble precipitation (from 1164.4 mm to 1272.5 mm), though individual model responses diverge widely. For instance, GFDL-ESM2M, which exhibits the highest projected rainfall anomaly (+40.0%, equivalent to +466 mm), simulates a dramatic +125.7% increase in streamflow, indicating strong nonlinear amplification likely driven by intensified rainfall events and saturation-excess runoff. Similarly, MIROC5, with a +27.0% increase in precipitation, yields a +58.3% discharge rise. In contrast, HadGEM2-ES and IPSL-CM5A-LR simulate notable rainfall deficits (−9.5% and −13.5%, respectively), which translate into reduced streamflows of −4.4% and −21.3% ( Table 4 , Figure 12 ). These disparities underscore the basin’s high rainfall-runoff sensitivity and the divergent hydrological pathways that arise from inter-model variability in precipitation projections. Overall, despite a modest ensemble mean increase in rainfall, the streamflow response is amplified in magnitude and variability, pointing to the potential for disproportionate hydrological extremes under future climate conditions—particularly in rainfall-dependent semi-humid systems like the Diaguiri Basin. These findings emphasize the importance of incorporating both ensemble means and model spread into future water resource planning and risk assessment frameworks.

The multi-model mean streamflow projections for the Diaguiri River Basin under the RCP4.5 and RCP8.5 scenarios highlight distinct temporal variability across three key sub-periods of the 21st century. Under the RCP4.5 scenario, the multi-model mean (MMM) streamflow increases by +24.3% in the near term (2021-2040), stabilizes at +20% by mid-century (2041-2070), and intensifies further to +30.7% in the late-century period (2071-2100). These hydrological shifts are consistent with the progressive increase in ensemble mean precipitation anomalies—rising from +6.72% to +12.16%—and reflect a coherent intensification of moisture availability under moderate radiative forcing. In contrast, under RCP8.5, the hydrological response exhibits a more erratic trajectory, shaped by highly divergent rainfall projections. While streamflows initially surge to +45.3% in the near term, they decline to +35% mid-century and further to +26.5% by century’s end, mirroring the weakening precipitation signal (+13.68% to +5.32%). This declining hydrological signal occurs despite some models projecting extreme increases in rainfall (e.g., GFDL-ESM2M: +62.7%), while others simulate severe drying (e.g., HadGEM2-ES: −31.7%). The broader inter-model spread in rainfall anomalies under RCP8.5 translates into greater hydrological uncertainty, with streamflow distributions exhibiting increased dispersion, asymmetry, and peakedness. Overall, the Diaguiri Basin’s strong rainfall-runoff coupling underscores the critical importance of accounting for precipitation variability in projecting future hydrological behavior. The contrasting patterns observed between scenarios point toward a wetter, yet more predictable future under RCP4.5, versus a more volatile and uncertain regime under RCP8.5—raising essential challenges for water resources planning and flood risk mitigation in the context of climate-driven non-stationarity.

The simulated monthly hydrographs of the Diaguiri River Basin, generated using the GR4J rainfall-runoff model driven by bias-corrected outputs from five CMIP5 climate models under RCP4.5 and RCP8.5 scenarios, indicate substantial shifts in the seasonal hydrological regime throughout the 21st century (Tables5&6). Both climate scenarios suggest a notable advancement in the onset of runoff, intensification of seasonal peak flows, and greater intra-annual variability. During the dry season (November-April), projected discharges remain low (below 1 m³/s), yet are slightly elevated compared to the historical baseline. For instance, January flows rise from 0.13 m³/s (observed) to 0.63 m³/s (projected), signaling possible changes in baseflow contributions or early-season precipitation inputs. Despite low mean flows, this period exhibits pronounced interannual variability, as shown by high skewness and kurtosis—particularly in May and June, where kurtosis exceeds 21 under RCP8.5, reflecting the likelihood of rare but intense rainfall-runoff events. The onset of significant runoff is projected to shift earlier in the hydrological year. In May, average flows reach 0.38 m³/s (RCP8.5) and 0.30 - 0.37 m³/s (RCP4.5), up from 0.27 m³/s observed, despite the flow regime being dominated by irregular and extreme events (CV > 2.0). This temporal advancement may reflect earlier rainfall onset and reduced soil moisture deficits, potentially altering catchment response thresholds.

Streamflow magnitudes during the peak season (July-September) show the most dramatic changes. Under RCP8.5, mean discharges in August and September rise to 25.86 m³/s and 35.05 m³/s respectively, corresponding to +54.8% and +13% increases over historical levels. RCP4.5 follows similar trends, albeit with more moderate increases. These changes are accompanied by heightened standard deviation and expanded interquartile ranges, indicating stronger interannual fluctuations. While coefficients of variation remain under 1—consistent with seasonal regimes—the rise in skewness (e.g., July: 1.06) and kurtosis (e.g., October: 3.51) reveals a higher frequency of extreme high-flow events, underscoring intensifying flood risks under future climate conditions. The recession period (October-December) also presents elevated flow levels. In October, for example, RCP8.5 projections indicate mean flows of 13.53 m³/s, compared to 11.57 m³/s historically (+16.9%). Statistical descriptors such as skewness and kurtosis remain high in these months, highlighting the persistence of asymmetric, heavy-tailed distributions and the continued risk of late-season hydrological extremes. Figure 15 and Figure 16 , which present monthly boxplots comparing historical and projected multi-model streamflow quantiles under RCP4.5 and RCP8.5, visually confirm the overall intensification of the seasonal hydrograph.

Overall, both emission scenarios forecast an intensification of runoff dynamics in the Diaguiri River Basin, characterized by heightened magnitudes, increased recurrence, and amplified variability of high-flow events. The RCP4.5 scenario suggests a moderately wetter hydrological regime accompanied by enhanced asymmetry and leptokurtic streamflow distributions, indicative of more frequent but moderate extremes. In contrast, RCP8.5 projects more severe hydrological extremes and greater inter-model and intra-annual variability, reflecting deeper distributional uncertainty. These projected transformations underscore a clear departure from historical hydrological stationarity and highlight the urgent need for flexible, adaptive water resource management strategies and robust flood risk governance tailored to the vulnerabilities of semi-humid tropical river systems such as the Diaguiri Basin.