Assessing Climate Change Impacts on Future Precipitation Using Random Forest Statistical Downscaling of CMIP6 HadGEM3 Projections in the Büyük Menderes Basin

Abstract

Climate change increasingly threatens the sustainability of regional water resources; therefore, robust station-scale precipitation projections are essential for basin-level planning. This study aims to develop and evaluate a hybrid, machine-learning-based statistical downscaling framework to generate monthly precipitation projections for the 21st century in the Büyük Menderes Basin, western Türkiye, using the HadGEM3-GC31-LL global climate model from the CMIP6. Monthly observations from 23 rainfall observation stations and ERA5 reanalysis predictors were employed to train station-specific Random Forest (RF) models, with optimal predictor sets identified through a multistage selection procedure (MPSP). Coarse-resolution general circulation model (GCM) fields were harmonized with ERA5 data using a three-stage inverse distance weighting (IDW), Delta, and Variance rescaling approach. The downscaled projections were bias-corrected using Quantile Delta Mapping (QDM) to maintain the climate-change signal. The RF models exhibited strong predictive skill across most stations, with test Nash–Sutcliffe Efficiency (NSE) values ranging from 0.45 to 0.81, RSR values from 0.43 to 0.74, and PBIAS values from −21.99% to +5.29%. Future projections indicate a basin-wide drying trend under both scenarios. Relative to the baseline, mean annual precipitation is projected to decrease by approximately 12.2, 19.6, and 33.7 mm in the near (2025–2050), mid (2051–2075), and late (2076–2099) periods under SSP2-4.5 (Shared Socioeconomic Pathway 2-4.5, a moderate greenhouse gas scenario). Under the high-emission SSP5-8.5 scenario, projected decreases are 25.2, 53.2, and 86.9 mm, respectively. Late-century reductions reach approximately 15–22% in several sub-basins. These findings indicate a substantial decline in future water availability and underscore the value of RF-based hybrid downscaling and trend-preserving bias correction for water resources planning in semi-arid Mediterranean basins.

1. Introduction

Climate change has progressively disrupted natural systems since the early stages of human civilization and has become a primary driver of global warming in recent decades. Greenhouse gas emissions associated with global economic growth and population expansion are a principal cause of this warming trend. As a result, the global hydrological cycle and the long-term sustainability of water resources face significant threats [1].

The spatial and temporal variability of water resources is highly responsive to changes in climatic conditions. This responsiveness highlights the necessity for comprehensive assessments of water resources in basins vulnerable to climate change [2]. Precipitation, a critical component of the hydrological cycle, has shown significant changes in recent decades due to climate forcing. Consequently, water resource planners and decision-makers in many regions have initiated adaptive strategies to address the projected impacts of climate change [3]. For Mediterranean-type climates, recent syntheses report increasing drought risk and water scarcity alongside shifts in precipitation regimes [4]. The effectiveness of these strategies depends on access to reliable future climate projections for the region of interest. General Circulation Models (GCMs) are currently the most advanced tools for generating climate projections of hydroclimatic variables at both global and regional scales.

The latest coordinated framework for global climate projections is the Coupled Model Intercomparison Project Phase 6 (CMIP6), organized under the World Climate Research Programme (WCRP). CMIP6 adopts a federated experimental design centered on a common set of baseline experiments (DECK) and historical simulations, complemented by a suite of endorsed Model Intercomparison Projects (MIPs) that address key scientific questions and provide a standardized basis for model evaluation and impact assessment [5]. In the Mediterranean–Türkiye hotspot context, CMIP6-based assessments have increasingly been used to characterize future hydroclimatic risks, and multi-model ensembles are often used for impact studies [6]. Future climate projections in CMIP6 are primarily generated through ScenarioMIP using Shared Socioeconomic Pathway (SSP) forcing trajectories such as SSP2-4.5 and SSP5-8.5, facilitating consistent scenario-based comparisons across models [7]. In this study, we use the CMIP6 model HadGEM3-GC31-LL (Met Office Hadley Centre), the low-resolution configuration of the HadGEM3-GC31 coupled model system [8]. The present study prioritizes development and demonstration of a reproducible station-scale hybrid SDS framework; therefore multi-model spread assessment is left for future work.

Although GCMs are valuable, their spatial resolution is insufficient to capture basin-scale climate processes. This limitation renders GCM outputs inadequate for hydrological applications that require detailed, fine-scale information [9,10]. To overcome this challenge, both dynamical and statistical downscaling methods have been developed to translate large-scale GCM simulations to local or basin scales [11]. Recent reviews highlight significant advances in statistical downscaling, particularly machine-learning approaches that better capture nonlinearity and complex terrain effects [12,13].

Dynamical downscaling utilizes physically based regional climate models to refine GCM projections; however, its use is limited by high computational requirements, model complexity, and strong dependence on GCM boundary conditions. In contrast, statistical downscaling (SDS) methods are more widely adopted due to their lower computational cost, operational simplicity, and robust predictive capabilities [3,14]. SDS approaches establish statistical transfer functions between large-scale GCM predictors and local scale predictands [15].

Despite their practical advantages, statistical downscaling methods also have limitations. They often assume that the statistical relationship between GCM-scale predictors and local climate remains approximately stable over time, a relationship that can be challenged by changing circulation and moisture regimes. Their performance may decline when applied to conditions outside the training range—such as rare events or situations strongly influenced by local processes (e.g., orographic precipitation). Finally, because SDS does not explicitly constrain solutions to physical laws, additional checks and post-processing may be required to maintain physical plausibility. Accordingly, careful predictor selection, scale harmonization, and trend-preserving post-processing are essential to improving the robustness of SDS-based projections.

The current literature shows that regression-based and ML-based models are frequently preferred due to their efficiency, ease of implementation, and strong predictive performance. Notable regression-based studies have utilized Multiple Linear Regression (MLR) [16,17,18,19,20] and Generalized Linear Models (GLMs) [21,22]. In the ML domain, Artificial Neural Networks (ANNs) have been widely and effectively applied to precipitation downscaling [23,24,25,26,27,28,29,30,31]. Support Vector Machines (SVMs) have also demonstrated strong performance in SDS applications [32,33,34,35,36].

Relevance Vector Machines (RVMs), which are sparse Bayesian analogs to SVMs, provide additional benefits such as reduced model complexity and fewer required parameters. Ghosh and Mujumdar [37] applied RVMs to downscale GCM simulations for basin-scale precipitation and streamflow, demonstrating clear performance advantages over SVMs. Subsequent research has further validated the effectiveness of RVMs for SDS applications [38,39,40]. Recent studies have also investigated hybrid and multi-method approaches, comparing or integrating Genetic Programming, ANN, SVM, and RVM techniques [41,42].

The Random Forest (RF) algorithm is another ML technique that has gained prominence due to its capacity to model nonlinear relationships and its resistance to overfitting, making it well-suited for complex hydroclimatic processes. Comparative studies indicate that RF outperforms other ML models in downscaling satellite-based precipitation datasets [43,44,45,46,47]. Recent assessments using the Coupled Model Intercomparison Project phase 6 (CMIP6), the latest generation of global climate models, confirm that RF often achieves superior downscaling performance compared to SVM across various climate regions [48]. RF has also shown strong skill in predicting daily and sub-daily precipitation from atmospheric variables [49].

In addition to GCM-based SDS, RF is widely used for downscaling precipitation and for integrating satellite data with reanalysis data. It frequently outperforms other ML methods and utilizes extra land-surface and topographic information. Recent explainable ML techniques, such as SHAP (SHapley Additive exPlanations), improve transparency by identifying main predictors and regional factors affecting rainfall [50,51].

Consistent with this evidence, recent CMIP6 case studies, including those from Türkiye, show that RF effectively downscales monthly precipitation. RF can be integrated with quantile-based bias-correction methods, either as a post-processing step or within hybrid correction frameworks [20,52,53].

Despite advances in CMIP6-based climate projections and the rising use of machine learning for statistical downscaling, several issues persist in applying these methods to basin-scale water resource management in Mediterranean semi-arid regions. First, station-scale monthly precipitation projections derived from CMIP6 simulations are still limited for many basins in Türkiye, particularly where complex topography and strong seasonal contrasts demand localized predictor–predictand relationships [6,20]. Second, many downscaling studies rely on either direct predictor–predictand mapping or a single-step bias adjustment. In contrast, fewer studies explicitly address the multi-stage inconsistency between coarse GCM fields and reanalysis-based predictors before model training. Third, while machine learning techniques like RF are increasingly popular, there is a notable lack of integrated hybrid frameworks that include (i) systematic, station-specific predictor selection, (ii) spatial reconciliation of GCM data with reanalysis at basin scales, and (iii) bias correction methods that preserve climate change signals in future projections.

To address these gaps, this study proposes a hybrid CMIP6-to-station downscaling framework for the Büyük Menderes Basin that integrates ERA5 predictors, station observations, and CMIP6 HadGEM3-GC31-LL outputs. The novelty of the approach lies in (1) implementing a station-specific multistage predictor-selection procedure to optimize RF predictors, (2) harmonizing coarse-resolution GCM fields with ERA5 through a three-stage IDW–Delta–Variance rescaling before downscaling, and (3) applying Quantile Delta Mapping (QDM) as a trend-preserving bias-correction step for SSP2-4.5 and SSP5-8.5 projections, thereby supporting consistent near-, mid-, and late-century assessments for basin-scale planning.

Decisions regarding water resources are typically made at the basin and station levels, but climate projections are often provided at coarse scales. The Büyük Menderes Basin in western Türkiye features a semi-arid, Mediterranean-type climate with marked seasonality and significant variability. Its water demand, mainly driven by irrigation, and recurring water stress heighten its sensitivity to climate-induced changes in rainfall. Consequently, accurate, station-level precipitation projections that retain the climate change signal are essential for developing resilient basin management strategies.

This study generates station-scale monthly precipitation projections for the Büyük Menderes Basin by integrating observations from 23 rain gauges, ERA5 predictors, and CMIP6 HadGEM3-GC31-LL simulations under SSP scenarios. It applies bias correction to reduce systematic errors in the projections.

The findings offer important insights into sustainable water resources management in semi-arid regions and provide a scientific basis for national climate adaptation strategies. This study constitutes a pioneering effort in Türkiye by presenting a hybrid SDS framework that integrates CMIP6 climate model outputs to assess regional hydrological impacts of climate change within the strategically significant Büyük Menderes Basin.

Accordingly, the main objectives of this study are to:

(i)

Develop and evaluate a station-specific RF-based statistical downscaling framework for monthly precipitation using ERA5 predictors and observations from 23 rain stations.

(ii)

Generate station-scale monthly precipitation projections from CMIP6 HadGEM3-GC31-LL under SSP2-4.5 and SSP5-8.5 for the near (2025–2050), mid (2051–2075), and late (2076–2099) periods.

(iii)

Quantify projected precipitation changes and their spatial variability across the basin to support water-resources planning in a semi-arid Mediterranean setting; and.

(iv)

Apply trend-preserving bias correction (Quantile Delta Mapping) to reduce systematic errors while retaining the climate-change signal in the downscaled projections.

2. Materials and Methods

2.1. Study Area

The Büyük Menderes Basin, the designated study area, is situated in western Türkiye within the Aegean Region as shown in Figure 1. Encompassing approximately 26,000 km², the Büyük Menderes Basin accounts for about 3.2% of Türkiye’s total land area and serves as a strategically significant water regime system [54]. The Büyük Menderes River, which extends 584 km from east to west, is the longest river discharging into the Aegean Sea. The basin is essential for supporting agricultural production, hydropower generation, and domestic water supply in the region.

The basin exhibits a typical Mediterranean climate, characterized by hot, dry summers and mild, wet winters. Annual mean temperatures range from 14 °C to 18 °C, with summer temperatures often exceeding 35 °C, particularly in lowland plains. In winter, average minimum temperatures range from 2 °C to 10 °C, and frost events are occasionally observed at higher elevations. The basin’s topography varies significantly, from rugged highlands in the east to fertile alluvial plains in the west. As a result, precipitation patterns are primarily influenced by frontal systems, while orographic precipitation is common in southern and mountainous regions [55]. According to the Köppen–Geiger climate classification, the Büyük Menderes Basin is categorized as ‘Csa’, indicating a hot-summer Mediterranean climate [56].

2.2. Data Sources

Monthly precipitation data from 23 meteorological stations within the basin were collected from official records for the period 1980–2014. During this interval, annual total precipitation ranged from 210 mm to 1305 mm. In recent years, rising temperatures and declining precipitation have led to significant water-supply challenges throughout the basin, particularly for agricultural irrigation.

This study utilized monthly precipitation data from 23 meteorological stations operated by the Turkish State Meteorological Service under the Ministry of Environment, Urbanization, and Climate Change. The locations of the stations are shown in Figure 1, and detailed information on each station is provided in

Table 1. Meteorological stations used in the Büyük Menderes Basin.

Number	Station Name	Station No	Latitude (°N)	Longitude (°E)	Elevation (m)
Station 1	Yatağan	17886	37.3395	28.1369	365
Station 2	Çine	18023	37.5947	28.0447	70
Station 3	Kavaklıdere	18022	37.4508	28.3621	973
Station 4	Kale	18301	37.4353	28.8608	1208
Station 5	Tavas	18065	37.5650	29.0706	992
Station 6	Söke	17881	37.7049	27.3827	75
Station 7	Aydın	17234	37.8402	27.8379	56
Station 8	Sultanhisar	17850	37.8843	28.1504	73
Station 9	Karacasu	18025	37.7194	28.6075	580
Station 10	Denizli	17237	37.7620	29.0921	425
Station 11	Nazilli	17860	37.9135	28.3437	84
Station 12	Kuyucak	18429	37.9022	28.4564	75
Station 13	Sarayköy	18068	37.9111	28.9081	194
Station 14	Çal	18067	38.0894	29.3953	817
Station 15	Dinar	17862	38.0597	30.1531	864
Station 16	Güney	17824	38.1515	29.0587	825
Station 17	Çivril	17825	38.2603	29.7120	824
Station 18	Eşme	17827	38.3961	28.9906	809
Station 19	Sivaslı	18071	38.4986	29.6953	990
Station 20	Hocalar	18292	38.5819	29.9697	1130
Station 21	Sandıklı	18002	38.4558	30.2313	1040
Station 22	Uşak	17188	38.6712	29.4040	919
Station 23	Banaz	18069	38.7525	29.7436	987

. The observation stations average rainfall was determined by calculating the mean of the total annual rainfall across the observation period.

The datasets used in this study include:

(i)

Historical precipitation observations from meteorological stations (1980–2014);

(ii)

Reanalysis datasets;

(iii)

GCM simulations from the CMIP6 framework representing both historical and future climate conditions.

The ERA5 reanalysis dataset, produced by the Copernicus Climate Change Service (C3S) of the European Centre for Medium-Range Weather Forecasts (ECMWF), was employed for the 1980–2014 period. In climate change research, datasets published by the ECMWF are frequently utilized in studies across the literature [57].

ERA5 reanalysis data have been widely employed in recent downscaling studies [58,59,60]. Furthermore, recent studies have shown that ERA5 performs well compared with other reanalysis products [61,62]. Data from 53 ERA5 grid points covering the study area were downloaded from the Copernicus Climate Data Store (https://cds.climate.copernicus.eu/datasets, accessed on 10 February 2024). The monthly ERA5 dataset has a spatial resolution of 0.25° × 0.25°, and the arrangement of the selected grids is shown in Figure 1.

Within the CMIP6 framework, this study employs a single GCM to demonstrate and evaluate the proposed hybrid statistical downscaling framework at the station scale. The model was selected primarily due to the completeness and internal consistency of the required variables and pressure levels, and the availability of continuous historical simulations together with SSP2-4.5/SSP5-8.5 projections from a harmonized source. In addition, a basic consistency check against ERA5 reanalysis was performed for the historical period to ensure a physically plausible baseline for the basin. To provide broader context on CMIP6 model performance for monthly precipitation over Türkiye, previous evaluation studies were also considered [63,64,65]. We acknowledge that using a single GCM does not quantify inter-model uncertainty across CMIP6; therefore, the projections should be interpreted as conditional on HadGEM3-GC31-LL. The primary contribution of this study is methodological: the proposed framework is model-agnostic and can be extended to multi-model CMIP6 ensembles in future work to explicitly assess inter-model uncertainty.

GCM generates future climate projections under various greenhouse gas emission pathways. In this study, the medium-emission scenario SSP2-4.5 (corresponding to 4.5 W/m² radiative forcing by 2100) and the high-emission scenario SSP5-8.5 (corresponding to 8.5 W/m² radiative forcing by 2100) were utilized. To resolve the spatial resolution mismatch between the GCM and ERA5 datasets, GCM grids were aligned with the ERA5 grid system using IDW and interpolation-based regridding.

Recent hydrological observations indicate a decline in precipitation across the basin. Consequently, the low-emission scenario SSP1-2.6 was excluded from the analysis. The historical period for monthly total precipitation from the GCM covers 1980–2014, while future SSP projections span 2025–2099. The CMIP6 GCM dataset utilized in this study was sourced from the Copernicus Climate Data Store (https://cds.climate.copernicus.eu/datasets). Details of the selected GCM are provided in

Table 2. The Data sets used in this study and their characteristics.

Data	Spatial Resolution	Number of Grid	Institution
HadGEM3-GC31-LL	1.875° × 1.25°	5	Met Office Hadley Centre
ERA5	0.25° × 0.25°	53	European Centre for Medium Range Weather Forecasts.

.

Regarding ensemble information, the historical, SSP2-4.5 and SSP5-8.5 simulations of HadGEM3-GC31-LL were taken from a single CMIP6 variant (r1i1p1f3) on the native grid (gn). Downscaling was applied to this single realization; no ensemble mean across multiple realizations was used.

This study utilized 12 atmospheric variables from four pressure levels (surface, 200 hPa, 500 hPa, and 850 hPa) from both the ERA5 reanalysis dataset and the selected CMIP6 HadGEM3-GC31-LL general circulation model dataset. The variables used in the analysis are listed in

Table 3. The common predictor variables selected from the ERA5 re-analysis and the GCM data set.

Number	Predictors	Description	Unit
1	pr	Monthly total precipitation	mm
2	tas	Near surface air temperature	°C
3	ps	Surface air pressure	hPa
4	zg200	200 hPa geopotential height	m
5	hur200	200 hPa relative humidity	%
6	ta200	200 hPa air temperature	°C
7	zg500	500 hPa geopotential height	m
8	hur500	500 hPa relative humidity	%
9	ta500	500 hPa air temperature	°C
10	zg850	850 hPa geopotential height	m
11	hur850	850 hPa relative humidity	%
12	ta850	850 hPa air temperature	°C

.

2.3. Methodology

Statistical downscaling methods are widely used to derive local-scale climate information from GCM outputs by modeling the statistical relationships between large-scale atmospheric predictors and local hydrometeorological observations. Once calibrated and validated, these relationships can be used in future GCM scenario simulations to generate localized precipitation projections [66]. In this study, SDS is implemented within a hybrid framework that combines reanalysis-based predictors with station observations and CMIP6 scenario data, enabling station-scale monthly precipitation projections suitable for basin-scale impact assessments.

This section provides sufficient methodological detail to ensure reproducibility. It outlines the process for screening station-specific predictors and details the setup and assessment of the Random Forest downscaling models. It also explains how GCM scenario predictors are prepared for use, including methods to minimize scale-related inconsistencies and the post-processing steps for the downscaled outputs used in scenario-based projections. Figure 2 clearly shows the workflow for this strategy, highlighting the sequence of steps and the flow of information between datasets. The subsequent subsections provide full methodological details, parameter settings, and performance metrics.

2.3.1. Multistage Predictor Selection Procedure

The performance of SDS models is strongly influenced by the relationship between selected atmospheric predictors and local meteorological variables. Consequently, the predictor selection process was conducted with particular care to identify the most appropriate multiple-predictor combinations for each station. Previous studies indicate that optimal predictors may vary according to the location of observation stations. Any predictor can be evaluated prior to the downscaling stage, but it should only be included if it demonstrates a statistically significant correlation with the observations [67].

Predictor selection was conducted independently for each observation station using a spatially weighted correlation-based method to identify the most representative ERA5 grids. This approach employs a scoring system that simultaneously considers both the distance between the station and the grids and the correlation coefficient. As a result, predictors derived from ERA5 grids with strong statistical and spatial representativeness were identified. Previous research has established that model performance is closely associated with the physical relevance of predictors to local hydrometeorological parameters and their degree of correlation [68,69,70].

To elucidate the multivariate relationships among predictor variables and optimize SDS model performance for each station, a multi-stage predictor selection procedure (MPSP) based on the All Possible Regression (APREG) approach was implemented. The procedure included the following steps: identifying ERA5 grids with high spatial representativeness, constructing predictor datasets for each station, conducting an exhaustive subset search using APREG with 5-fold cross-validation, performing multi-criteria model scoring, and executing a sensitivity analysis. The development of these steps was guided by the principles of physical consistency, generalization capability, and parsimony as emphasized in the literature [71,72].

The most representative ERA5 grids for each station were determined by jointly evaluating the correlation coefficient (R) between the grid and station time series and the corresponding distance, to compute a spatially weighted correlation. This method identified the four most representative grids for each station. The average of these four grid time series was then used to generate predictor time series for all ERA5 variables listed in Table 3. This process improved representativeness and minimized noise.

In constructing predictor subsets, the APREG approach, which has demonstrated effectiveness in recent SDS studies, was applied [24,67]. For each station, all possible combinations of 1 to 12 predictors were generated. The primary evaluation metric was root mean square error (RMSE) using 5-fold cross-validation, with additional metrics including adjusted coefficient of determination (Adj. R²), Mallows’ Cp, and the Bayesian Information Criterion (BIC). In accordance with parsimony principles, these penalty-based metrics were weighted under alternative scenarios and analyzed through sensitivity analysis to compute a combined score coefficient. This methodology mitigated overfitting through cross-validation and simultaneously optimized model training performance, physical consistency, and generalization ability. Penalty-based criteria such as Cp and BIC were used to balance model simplicity and predictive performance. Sensitivity analysis quantified the impact of alternative weighting scenarios on model selection, ensuring the identification of stable predictor combinations [73,74,75].

The MPSP analysis was conducted for each station using precipitation data from 23 meteorological stations and the ERA5 dataset. The optimal predictor combination for each station was identified and recommended for subsequent model training. For example,

Table 4. Results of the multistage predictor selection procedure for Banaz station.

Predictor Number	pr	tas	ps	zg200	hur200	ta200	zg500	hur500	ta500	zg850	hur850	ta850	R2Adj	RMSE	Cp	BIC	Score
1	√												0.734	19.38	38.57	2487	0.00
2	√									√			0.739	19.04	31.19	2484	0.35
3	√								√		√		0.750	18.70	14.25	2472	0.69
4	√				√		√				√		0.758	18.31	0.60	2462	0.99
5	√	√			√				√		√		0.759	18.35	0.97	2467	0.95
6	√	√	√	√	√						√		0.758	18.37	3.20	2473	0.88
7	√	√	√		√					√	√	√	0.758	18.36	3.99	2478	0.83
8	√	√	√	√	√			√	√		√		0.758	18.39	5.41	2483	0.76
9	√		√	√		√	√	√	√	√	√		0.757	18.38	7.79	2490	0.67
10	√	√	√	√			√	√	√	√	√	√	0.757	18.47	9.44	2495	0.55
11	√	√	√	√	√		√	√	√	√	√	√	0.757	18.59	9.00	2499	0.45
12	√	√	√	√	√	√	√	√	√	√	√	√	0.757	18.73	11.00	2505	0.00

presents the MPSP results for the Banaz station, where the four-variable predictor combination achieved the highest score coefficient, whereas combinations with more predictors yielded lower scores. Therefore, the MPSP approach facilitated the use of simple yet effective predictor combinations in model training, avoiding unnecessary complexity.

2.3.2. GCM–ERA5 Datasets Harmonization Using IDW–Delta–Variance Approach

GCM typically produce outputs at coarse spatial resolutions, creating challenges for hydrological and climate-impact studies conducted at the basin scale. The mismatch between GCM resolutions and those of observational or reanalysis datasets impedes meaningful comparisons and model integration. Therefore, harmonizing GCM data with higher-resolution reanalysis datasets is a critical preprocessing requirement [76,77].

A three-stage regridding framework was implemented to adapt GCM outputs to the 0.25° × 0.25° spatial resolution of the ERA5 dataset. This procedure comprises the following steps:

1.

Inverse Distance Weighting: During the first stage, GCM fields were interpolated onto ERA5 grid points using the IDW method, which assigns weights inversely proportional to the distance between source and target grid cells. This approach is a fundamental method for spatially harmonizing datasets with different resolutions. The IDW formulation is provided in Equation (1) [78]:

$$\mathrm{ŷ}\left(x_0\right)={\displaystyle \frac{{\sum }_{i=1}^N\frac{y_i}{{d\left(x_0,x_i\right)}^p}}{{\sum }_{i=1}^N\frac{1}{{d\left(x_0,x_i\right)}^p}}}$$

(1)

In this formulation, ŷ represents the estimated value in the target grid, $y_i$ denotes the value in the source grid, $d\left(x_0,x_i\right)$ is the distance in kilometers between the target and source grids, and p is the weight parameter.

2.

Delta (Anomaly) Adjustment: To preserve projected climate change signals within the context of ERA5 baseline climatology, the delta (anomaly) method was applied. In this step, GCM future-period deviations were calculated relative to the model’s historical mean climate, and these anomalies were then added to the ERA5 climatological normals [79]. The formulation for Delta Anomaly is given in Equation (2):

$$\begin{gathered} Y_{anom}\left(t\right)=Y_{GCM}\left(t\right)-{Ȳ}_{GCM,ref} \\ {Y}_{regrid}\left(t\right)={Ȳ}_{ERA5,ref}-Y_{anom}\left(t\right) \end{gathered}$$

(2)

In this equation, $Y_{anom}$ denotes the GCM anomaly value, $Y_{GCM}$ is the GCM raw data, ${Ȳ}_{GCM,ref}$ is the reference period (1980–2014) average of the GCM, $Y_{regrid}$ is the GCM value adapted to the ERA5 scale, and ${Ȳ}_{ERA5,ref}$ is the ERA5 reference period average.

3.

Variance Scaling (Variance Post-processing): In the final stage, the variability of the regridded GCM series was adjusted to match the statistical dispersion of the ERA5 series. This correction reduces systematic variance differences and improves statistical compatibility between the two datasets [76]. The details are presented in Equation (3).

$${Y\prime }_{regrid}\left(t\right)=\mu +\left(Y_{regrid}-\mu \right)\times {\displaystyle \frac{{\sigma }_{ERA5}}{{\sigma }_{GCM}}}$$

(3)

where ${Y\prime }_{regrid}$ represents the variance-corrected regridded data, $\mu$ is the mean of the regridded series, ${\sigma }_{ERA5}$ is the standard deviation of ERA5, and ${\sigma }_{GCM}$ is the standard deviation of the GCM.

The combined IDW–Delta–Variance procedure provides a comprehensive preprocessing framework for harmonizing GCM projections with the high-resolution ERA5 dataset before integration into the RF–based statistical downscaling model. This approach resolves the spatial-resolution mismatch between GCM and ERA5 and preserves future climate change signals in GCM dataset through anomaly and variance adjustments.

This method advances beyond conventional interpolation techniques by simultaneously achieving spatial consistency, maintaining projected climate signals, and enhancing statistical fidelity [29,80,81]. As a result, the rescaled GCM dataset were effectively integrated into station-based Random Forest models, leading to substantial improvements in both the reliability and spatial representativeness of future precipitation projections.

2.3.3. Random Forest Method

RF is a powerful and flexible ensemble learning method that constructs a large number of decision trees on bootstrap samples and aggregates their outputs to generate robust predictions. As an ensemble-based, nonparametric machine learning approach, RF is particularly effective at capturing nonlinear relationships and complex high-order interactions among predictors [82]. Within the RF framework, each tree is trained on a randomly drawn subset of the training data, while a randomly selected subset of predictor variables is considered at each node. Each tree learns independently and produces an individual prediction; for regression tasks, the ensemble output is obtained by averaging the predictions across all trees [83].

Randomly selecting both training samples and predictor subsets reduces correlation among trees and, consequently, decreases generalization error. As demonstrated by Breiman [82], the generalization error of a random forest converges toward a limiting value as the number of trees increases. A key advantage of RF is that it does not rely on assumptions regarding the distribution of input variables and is insensitive to variable scaling. The method can therefore produce unbiased or near-unbiased predictions, while effectively handling datasets containing outliers or missing values [84].

ANNs are widely used in statistical downscaling because they can model complex nonlinear relationships between large-scale predictors and local precipitation. Nevertheless, the effectiveness of ANNs depends heavily on network architecture and training methods, often requiring careful predictor normalization and risking overfitting when the model has many parameters relative to the training data. Conversely, RF uses bootstrap aggregation and random predictor subsampling to reduce variance and improve generalization. It is less affected by predictor scaling, provides an internal out-of-bag error estimate, and includes variable metrics that aid physical interpretation. These features make RF especially suitable for station-scale precipitation downscaling, as it handles multiple correlated predictors and nonlinear, intermittent precipitation patterns effectively.

Although conceptually straightforward, RF frequently surpasses conventional classifiers and regressors such as discriminant analysis, support vector machines, and artificial neural networks, primarily due to its robustness against overfitting. Furthermore, RF implementation is straightforward, requiring specification of only two primary hyperparameters: the number of trees in the ensemble (ntree) and the number of predictors randomly sampled at each node (mtry). Model performance typically exhibits low sensitivity to these parameter choices [82], which supports the practicality and stability of RF for large-scale climate applications.

In the context of statistical downscaling, the RF algorithm has several notable advantages. Precipitation is influenced by numerous atmospheric variables, and RF can efficiently process high-dimensional predictor sets without requiring explicit feature selection [85]. These properties have contributed to the growing adoption of RF in climate and hydrological research for both regression and classification tasks [86,87,88,89,90,91]. The operational steps of the RF algorithm used in this study are summarized as follows:

1.

Determination of the number of trees (ntree): A predefined number of decision trees is specified. For each tree, a bootstrap sample is generated by sampling with replacement from the training data. Observations not included in this sample form the out-of-bag (OOB) set for that tree.

2.

Tree construction using random predictor subsets: For each bootstrap sample, an unpruned regression tree is grown. At each node, mtry predictor variables are randomly selected from the full set, and the best split among these candidates is chosen. This process ensures diversity across trees by combining randomness in both data sampling and predictor selection.

3.

Prediction aggregation: For new observations, predictions from all trees are aggregated by averaging in regression tasks or by majority voting in classification tasks to generate the final RF estimate.

4.

OOB error estimation: Each tree is used to predict the OOB observations corresponding to it. Once all OOB predictions are obtained, the out-of-bag root mean square error (OOB-RMSE) is calculated. This provides an unbiased and efficient estimate of the model’s generalization error without requiring a separate validation dataset.

5.

Model selection: The RF configuration yielding the lowest prediction error during the iterative training process is selected as the optimal model for downscaling [82]

Overall, the RF approach offers a robust, computationally efficient, and flexible framework for statistical downscaling, making it suitable for modeling precipitation variability and future climate projections.

2.3.4. Performance Metrics

To assess the performance of the RF-based statistical downscaling models developed for each station, several performance metrics were calculated. These metrics include RMSE, Nash–Sutcliffe Efficiency coefficient ($NSE$), adjusted coefficient of determination (${AdjR}^2$), percent bias ($PBIAS$), and the ratio of root mean square error to the standard deviation of observations ($RSR$), which has become increasingly prominent in hydrology and climate modeling studies [92]. While NSE was used to evaluate predictive skill, Adj R² was employed to account for the effect of predictor dimensionality in the regression-based downscaling framework.

In addition, OOB-RMSE is directly computed from out-of-bag predictions generated during RF training, providing an internal estimate of generalization errors. The equations for these metrics are provided in Equations (4)–(8), respectively.

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}} \sum _{i=1}^n{\left(Y_i^{Obs}-Y_i^{Mod}\right)}^2}$$

(4)

$$NSE=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(Y_i^{Obs}-Y_i^{Mod}\right)}^2}{{\sum }_{i=1}^n{\left(Y_i^{Obs}-{Ȳ}^{Obs}\right)}^2}}$$

(5)

$${AdjR}^2=1-(1-R^2){\displaystyle \frac{n-1}{n-P-1}}$$

(6)

$$PBIAS=100\times {\displaystyle \frac{{\sum }_{i=1}^N\left(Y_i^{Mod}-Y_i^{Obs}\right)}{{\sum }_{i=1}^n{Y}_i^{Obs}}}$$

(7)

$$RSR={\displaystyle \frac{RMSE}{{\sigma }_{Obs}}}$$

(8)

In the equations, $Y_i^{Obs}$ indicates the observed values, $Y_i^{Mod}$ is the model output, ${\mathrm{Ȳ}}^{Obs}$ denotes the mean of the observed data, ${\sigma }_{Obs}$ represents the standard deviation of the observed data, where $n$ is the number of observations and $P$ is the number of predictors used in the station-specific RF model. Lower RMSE and RSR values indicate better performance, while NSE values close to 1 signify higher prediction accuracy. PBIAS values near 0 suggest minimal bias. Positive PBIAS indicates overall overestimation, whereas negative values indicate underestimation.

2.3.5. Bias Correction

Outputs from GCM and SDS models often exhibit systematic distributional biases due to spatial resolution limitations, structural deficiencies in physically based climate models, methodological choices in optimization and downscaling. When uncorrected, especially in highly skewed and variable fields such as precipitation, these biases can result in an inaccurate representation of mean precipitation and an erroneous estimation of the frequency and intensity of extreme events. Such inaccuracies undermine the reliability of hydrological impact assessments.

Accordingly, quantile mapping techniques are widely adopted to correct systematic distributional biases in climate model precipitation outputs. In the present hybrid framework, bias correction is applied after statistical downscaling, i.e., as a post-processing step on station-scale precipitation projections, rather than bias-correcting coarse-grid GCM precipitation prior to downscaling. This sequencing is adopted for two main reasons. First, the RF downscaling models are calibrated solely using ERA5 predictors and station observations, and GCM outputs are used only at the projection stage; consequently, applying a bias-correction procedure to GCM precipitation before downscaling is not part of the model calibration pathway and does not necessarily improve the distribution of station-scale precipitation. Second, precipitation biases are scale-dependent: correcting biases at the native GCM grid does not guarantee unbiased station-scale values after downscaling, and the downscaling step may introduce or amplify distributional inconsistencies, which could necessitate an additional correction at the target scale. Therefore, QDM is applied once to the downscaled station precipitation to correct residual distributional biases directly at the impact-relevant scale, while preserving the projected change signal.

Traditional quantile mapping techniques effectively correct discrepancies during the historical period relative to observations; however, they can artificially distort long-term trends in future model projections [93]. To address this limitation, the present study employs the Quantile Delta Mapping (QDM) method, an advanced quantile mapping approach that preserves the projected change signal of GCM scenarios when correcting biases in station-scale future precipitation projections.

Within the QDM framework, the quantile of a model-projected future value is first determined from the cumulative distribution function (CDF) of the future model period. The corresponding historical model value is then identified using the inverse CDF of the historical model distribution. The ratio of these two values quantifies the relative change, or delta, between the historical and projected model climates.

This quantile is then mapped onto the observational distribution using the inverse CDF of the historical observations. The bias-corrected future model value is obtained by multiplying the observational quantile value by the delta factor [93]. Through this process, QDM preserves the climate change signal present in GCM and corrects systematic biases. The method is expressed as follows Equations (9)–(12):

$$q_{m,f}\left(t\right)= F_{m,f}^t\left[x_{m,f}\left(t\right)\right], q_{m,f}\left(t\right)\in \left\{\mathrm{0,1}\right\}$$

(9)

$${∆}_m\left(t\right)={\displaystyle \frac{x_{m,f}\left(t\right)}{F_{m,h}^{-1}\left[q_{m,f}\left(t\right)\right] }}$$

(10)

$$x_{o,h}\left(t\right)=F_{o,h}^{-1}\left[q_{m,f}\left(t\right)\right]$$

(11)

$${x\prime }_{m,f}\left(t\right)=x_{o,h} \left(t\right)\times {∆}_m\left(t\right)$$

(12)

In these formulations, $x_{m,f}$ denotes the model projection, $F_{m,f}$ represents the CDF of the future period model, $q_{m,f}$ is the quantile of the future model value, $F_{m,h}^{-1}$ is the inverse CDF of the historical model distribution,${∆}_m\left(t\right)$ shows the relative change between the projected and historical model values, $x_{o,h}$ is the quantile value from the observational CDF, $F_{o,h}^{-1}$ is the inverse CDF of the historical observational distribution, and ${x\prime }_{m,f}$ denotes the bias-corrected future projection.

Owing to its ability to preserve trends, QDM is regarded as one of the most reliable bias-correction techniques for climate change impact studies [76]. It has been widely implemented in recent climate and hydrological modeling research [94,95]. Comparative evaluations of bias-correction techniques further demonstrate that QDM consistently outperforms alternative methods [96,97,98].

2.3.6. Future Precipitation Projections Under SSP Scenarios

Station-specific RF models were developed using ERA5 predictors and monthly precipitation observations. Following the split shown in Figure 3, RF hyperparameters were optimized in the training period and finalized based on out-of-sample performance in the test period.

The finalized RF models were then applied to GCM simulations to project station-scale precipitation under the SSP2-4.5 and SSP5-8.5 scenarios. For each station, the predictor sets identified by the MPSP were retained to ensure consistency between the historical calibration and scenario applications. Since the RF models were trained using ERA5-based predictors, the GCM fields were first adjusted to align with ERA5 to reduce scale discrepancies between the reanalysis and GCM datasets. Specifically, GCM grids were harmonized with ERA5 grids using the three-stage IDW–Delta–Variance rescaling procedure. The harmonized scenario predictors were then used as inputs to the trained RF models to produce raw downscaled monthly precipitation projections for the entire scenario period (2025–2099).

To correct residual distributional biases while preserving the climate-change signal, the raw projections were bias-corrected using QDM. Although the SSP simulations begin in 2015, results are summarized for three standard future windows—near future (2025–2050), mid-century (2051–2075), and late century (2076–2099)—to highlight changes relative to the historical baseline and to ensure consistency with commonly used reporting periods in climate-impact studies.

3. Results

3.1. Results of Statistical Downscaling Model

During model training, precipitation data from observation stations served as the dependent variable, while multiple atmospheric variables from the ERA5 reanalysis dataset were used as predictors. This approach enabled the capture of multivariate and nonlinear relationships between large-scale climate variables and station-scale precipitation observations. Separate RF-based downscaling models were developed for each station using MATLAB R2025b. The workflow for this strategy is illustrated in Figure 2.

The performance of the RF algorithm depends significantly on the specification of its hyperparameters. Three key hyperparameters were optimized individually for each station using Bayesian Optimization: the number of trees (ntree), which was varied from 50 to 500 to balance flexibility and generalization; the minimum leaf size (MinLeafSize), adjusted between 1 and 20 to mitigate overfitting; and the number of predictor variables at each split (mtry), explored from 1 to p to maintain model diversity and an appropriate bias–variance balance. After dynamic hyperparameter optimization, the RF models were trained for each station.

3.2. Evaluation of Statistical Downscaling Models

The observational and ERA5 reanalysis datasets were divided into two periods for training and testing the downscaling models: 50% for training (1980–1997) and 50% for testing (1998–2014). A key advantage of the RF algorithm is its capacity to internally evaluate model performance using OOB samples. In this study, model performance was assessed using both OOB samples during training and an independent test dataset. OOB samples provide an estimate of internal accuracy by evaluating the model on data not used during training, while the independent test set measures external accuracy by assessing performance on previously unseen data.

Accordingly, a two-step validation approach was employed: internal validation using OOB-RMSE and external validation using test-period performance metrics. This dual validation strategy enhances the scientific rigor of the analysis by demonstrating both the model’s resistance to overfitting and its predictive performance.

During the test period, RMSE values varied between 13.30 and 37.48 mm, while RSR efficiency for the SDS models ranged from 0.43 to 0.74. NSE scores fell between 0.45 and 0.81, and PBIAS values ranged from 5.29% to −21.99%. Additionally, OOB-RMSE values computed by the RF algorithm during model training ranged from 14.95 to 29.48.

Among the 23 observational stations evaluated, test-period NSE values of the RF models were classified as “good” to “very good” for all except five stations, while RSR values received similar ratings for all except four stations. PBIAS values were also assessed as “good” to “very good” for all but one station. These evaluations are based on performance classification criteria widely used in hydrological and climate modeling studies [99,100].

Test period performance metrics and internal validation results from the RF algorithm are presented in

Table 5. Results of performances of RF models trained for the observation stations.

Number	Station Name	Test RMSE (mm)	RSR	NSE	Adj R2	PBIAS (%)	OOB RMSE (mm)
Station 1	Yatağan	26.94	0.45	0.80	0.80	−4.78	26.83
Station 2	Çine	21.41	0.52	0.73	0.74	−11.42	23.73
Station 3	Kavaklıdere	25.25	0.46	0.79	0.79	−3.70	24.22
Station 4	Kale	34.70	0.62	0.61	0.68	−21.99	26.02
Station 5	Tavas	14.80	0.47	0.78	0.79	−6.85	17.55
Station 6	Söke	37.48	0.45	0.80	0.80	0.20	29.48
Station 7	Aydın	30.49	0.52	0.73	0.72	−3.15	24.71
Station 8	Sultanhisar	25.38	0.43	0.81	0.81	−7.11	23.05
Station 9	Karacasu	27.55	0.52	0.72	0.73	−5.26	26.97
Station 10	Denizli	23.41	0.49	0.76	0.79	−10.94	22.69
Station 11	Nazilli	26.74	0.52	0.73	0.73	−0.53	25.22
Station 12	Kuyucak	22.03	0.51	0.74	0.74	−4.98	18.04
Station 13	Sarayköy	13.30	0.46	0.79	0.79	−4.59	16.76
Station 14	Çal	23.21	0.62	0.62	0.62	−10.16	16.75
Station 15	Dinar	17.86	0.57	0.67	0.68	−11.38	16.51
Station 16	Güney	19.95	0.52	0.73	0.73	0.84	18.22
Station 17	Çivril	18.09	0.58	0.67	0.67	−4.12	16.66
Station 18	Eşme	19.37	0.59	0.65	0.66	−10.73	16.26
Station 19	Sivaslı	17.32	0.52	0.73	0.74	−7.66	18.42
Station 20	Hocalar	18.54	0.74	0.45	0.50	2.95	15.60
Station 21	Sandıklı	16.65	0.71	0.49	0.49	5.29	14.95
Station 22	Uşak	20.52	0.53	0.71	0.72	−9.22	16.78
Station 23	Banaz	18.92	0.51	0.74	0.75	−9.97	20.62

. An example of training and test performance for the Yatağan station is shown in Figure 3.

Comparison of OOB-RMSE and test RMSE values across stations indicates that the RF method generally performs consistently at most locations. The close agreement between OOB-RMSE and test RMSE suggests that the model does not overfit during training [82]. However, larger discrepancies observed at stations such as Station 4, Station 6, and Station 14 may result from local data quality issues, station-grid inconsistencies, or climatic anomalies. These findings emphasize the importance of station-specific evaluations and the need to account for spatial heterogeneity in SDS applications.

3.3. Downscaling of Future Precipitation and Bias Correction

Accurately estimating future climate change impacts is critical for effective planning and sustainable management of regional water resources. The limited spatial resolution of GCM, which serve as primary data sources for climate change assessments, restricts their direct application at basin or station scales. To overcome this constraint, the spatial resolution of GCM data was improved to align with that of the ERA5 dataset during the development of RF-based statistical downscaling models, as previously described.

Two scenarios from the HadGEM3-GC31-LL climate model within the CMIP6 framework were employed: SSP2-4.5, representing a moderate emissions pathway with mitigation measures such as energy efficiency improvements and carbon reduction strategies, and SSP5-8.5, reflecting a high-emission future in which such measures are absent and greenhouse gas emissions increase. Following the alignment of GCM output and ERA5 reanalysis data resolutions, statistical downscaling was performed using RF models trained on optimal predictor combinations for each of the 23 stations.

Extensive research indicates that downscaled results frequently exhibit significant biases, primarily due to model parameterizations and unresolved geophysical processes [19,101,102]. To address these biases, the QDM method introduced by Cannon et al. [93] was applied. Theoretical formulations and implementation steps for QDM have been detailed in previous sections. QDM was chosen for its proven ability to preserve the climate change signal in future projections.

The bias-correction procedure was applied to the downscaled future precipitation datasets, resulting in a substantial reduction in systematic biases. CDFs were generated for observed precipitation, uncorrected projections, and bias-corrected projections at all 23 stations. Figure 4 presents the CDFs for the Yatağan station (Station 1), along with the corresponding PBIAS values before and after bias correction. Due to space constraints, only the results for Yatağan are shown.

Following bias correction, future precipitation projections from the SSP2-4.5 and SSP5-8.5 scenarios of the HadGEM3-GC31-LL model were evaluated at the station scale for three periods: near future (2025–2050), mid-future (2051–2075), and far future (2076–2099). For all stations, mean observed annual precipitation, mean projected annual precipitation, and projected changes are summarized in

Table 6. Estimated annual total precipitation and changes at stations under the SSP2-4.5 scenario.

Number	Station Name	Observed (mm)	Near Future (mm)	Middle Future (mm)	Far Future (mm)	Near Change (mm)	Middle Change (mm)	Far Change (mm)	Near Change (%)	Middle Change (%)	Far Change (%)
Station 1	Yatağan	661.5	647.9	647.6	611.4	−13.7	−13.9	−50.2	−2.1	−2.1	−7.6
Station 2	Çine	547.3	530.1	537.0	501.7	−17.2	−10.4	−45.6	−3.15	−1.89	−8.33
Station 3	Kavaklıdere	729.0	693.7	672.5	647.9	−35.4	−56.6	−81.1	−4.8	−7.8	−11.1
Station 4	Kale	767.4	756.0	737.5	702.9	−11.4	−29.8	−64.5	−1.5	−3.9	−8.4
Station 5	Tavas	495.4	499.5	492.3	460.9	4.1	−3.1	−34.6	0.8	−0.6	−7.0
Station 6	Söke	834.6	855.8	847.8	796.6	21.1	13.2	−38.1	2.5	1.6	−4.6
Station 7	Aydın	623.7	613.8	614.9	581.2	−10.0	−8.8	−42.6	−1.6	−1.4	−6.8
Station 8	Sultanhisar	593.2	577.0	581.1	555.0	−16.2	−12.1	−38.2	−2.7	−2.0	−6.4
Station 9	Karacasu	605.0	589.3	578.3	575.2	−15.7	−26.7	−29.8	−2.59	−4.42	−4.92
Station 10	Denizli	562.9	555.7	539.7	526.2	−7.1	−23.2	−36.6	−1.27	−4.12	−6.51
Station 11	Nazilli	562.0	532.3	518.9	517.3	−29.7	−43.2	−44.7	−5.3	−7.7	−8.0
Station 12	Kuyucak	481.4	459.9	458.4	444.1	−21.5	−23.0	−37.3	−4.5	−4.8	−7.8
Station 13	Sarayköy	372.9	356.8	346.1	346.2	−16.2	−26.8	−26.8	−4.3	−7.2	−7.2
Station 14	Çal	478.7	467.5	456.4	467.2	−11.2	−22.3	−11.5	−2.3	−4.7	−2.4
Station 15	Dinar	443.1	433.1	431.9	432.1	−10.0	−11.2	−11.0	−2.3	−2.5	−2.5
Station 16	Güney	502.4	492.2	471.8	483.2	−10.2	−30.7	−19.3	−2.0	−6.1	−3.8
Station 17	Çivril	437.8	422.9	420.6	408.5	−14.9	−17.1	−29.2	−3.4	−3.9	−6.7
Station 18	Eşme	456.4	453.1	445.6	444.4	−3.3	−10.8	−12.1	−0.7	−2.4	−2.6
Station 19	Sivaslı	494.5	473.5	473.9	468.6	−21.0	−20.6	−25.9	−4.2	−4.2	−5.2
Station 20	Hocalar	433.0	433.1	412.1	397.6	0.0	−20.9	−35.4	0.0	−4.8	−8.2
Station 21	Sandıklı	406.5	408.6	398.5	381.9	2.1	−8.0	−24.6	0.5	−2.0	−6.1
Station 22	Uşak	525.9	507.7	505.7	506.2	−18.2	−20.2	−19.7	−3.5	−3.8	−3.7
Station 23	Banaz	526.9	502.7	501.9	510.8	−24.2	−25.0	−16.1	−4.6	−4.7	−3.1

and

Table 7. Annual total precipitation estimates and changes at stations under the SSP5-8.5 scenario.

Number	Station Name	Observed (mm)	Near Future (mm)	Middle Future (mm)	Far Future (mm)	Near Change (mm)	Middle Change (mm)	Far Change (mm)	Near Change (%)	Middle Change (%)	Far Change (%)
Station 1	Yatağan	661.5	632.8	567.0	549.5	−28.8	−94.6	−112.0	−4.4	−14.3	−16.9
Station 2	Çine	547.3	519.7	487.4	470.9	−27.6	−59.9	−76.4	−5.05	−10.95	−13.96
Station 3	Kavaklıdere	729.0	712.7	634.8	615.8	−16.4	−94.3	−113.2	−2.2	−12.9	−15.5
Station 4	Kale	767.4	748.9	701.4	681.1	−18.5	−66.0	−86.3	−2.4	−8.6	−11.2
Station 5	Tavas	495.4	486.9	468.3	432.4	−8.5	−27.1	−63.0	−1.7	−5.5	−12.7
Station 6	Söke	834.6	813.4	803.2	710.6	−21.3	−31.4	−124.0	−2.6	−3.8	−14.9
Station 7	Aydın	623.7	589.1	540.0	510.7	−34.6	−83.7	−113.0	−5.6	−13.4	−18.1
Station 8	Sultanhisar	593.2	559.9	518.6	490.8	−33.4	−74.7	−102.4	−5.6	−12.6	−17.3
Station 9	Karacasu	605.0	596.0	536.6	505.3	−9.0	−68.4	−99.7	−1.49	−11.30	−16.48
Station 10	Denizli	562.9	560.6	506.8	474.5	−2.2	−56.1	−88.4	−0.39	−9.96	−15.71
Station 11	Nazilli	562.0	514.2	486.2	450.1	−47.8	−75.9	−111.9	−8.5	−13.5	−19.9
Station 12	Kuyucak	481.4	441.9	419.8	396.3	−39.6	−61.6	−85.1	−8.2	−12.8	−17.7
Station 13	Sarayköy	372.9	345.2	322.9	302.5	−27.7	−50.1	−70.5	−7.4	−13.4	−18.9
Station 14	Çal	478.7	446.2	441.0	377.5	−32.5	−37.8	−101.2	−6.8	−7.9	−21.1
Station 15	Dinar	443.1	428.6	417.1	395.0	−14.5	−26.0	−48.1	−3.3	−5.9	−10.9
Station 16	Güney	502.4	463.1	456.2	395.3	−39.3	−46.3	−107.2	−7.8	−9.2	−21.3
Station 17	Çivril	437.8	412.5	397.1	368.1	−25.3	−40.7	−69.6	−5.8	−9.3	−15.9
Station 18	Eşme	456.4	425.7	422.9	386.9	−30.7	−33.5	−69.6	−6.7	−7.3	−15.2
Station 19	Sivaslı	494.5	461.4	444.5	414.9	−33.1	−50.0	−79.6	−6.7	−10.1	−16.1
Station 20	Hocalar	433.0	414.0	397.8	362.0	−19.1	−35.3	−71.0	−4.4	−8.1	−16.4
Station 21	Sandıklı	406.5	402.4	379.2	354.6	−4.1	−27.3	−51.9	−1.0	−6.7	−12.8
Station 22	Uşak	525.9	489.3	483.9	455.1	−36.6	−42.0	−70.8	−7.0	−8.0	−13.5
Station 23	Banaz	526.9	497.3	486.2	444.2	−29.6	−40.7	−82.8	−5.6	−7.7	−15.7

.

Compared to the observation period, the SSP2-4.5 scenario projects basin-wide declines in annual total precipitation, averaging 12.2 mm/year in the near future, 19.6 mm/year in the mid-future, and 33.7 mm/year in the far future. Four stations display increases in annual precipitation in the near future; however, only the increase at Söke station is statistically significant. The remaining nineteen stations exhibit declines, with particularly notable reductions (greater than 20 mm/year) at Kavaklıdere, Nazilli, Kuyucak, Sivaslı, and Banaz. In the mid-future, all stations except Söke experience decreases in annual total precipitation, with the largest reduction (56.6 mm/year) at Kavaklıdere. Far-future projections indicate substantial reductions across all stations, with decreases reaching up to 81.1 mm/year. Detailed results are provided in Table 6.

Under the SSP5-8.5 scenario, reductions in mean annual precipitation are more pronounced relative to the observation period: 25.2 mm in the near future, 53.2 mm in the mid-future, and 86.9 mm in the far future. The largest near-future decline (47.8 mm/year) is projected for Nazilli station. Mid-future reductions intensify, reaching up to 94.6 mm/year at certain stations. In the far-future period, decreases approach 124 mm/year across the basin. Comprehensive results for SSP5-8.5 are presented in Table 7.

In summary, the bias-corrected RF-based downscaling results demonstrate a consistent decline in precipitation across the Büyük Menderes Basin under both emission scenarios, with a more pronounced trend under the high-emission SSP5-8.5 pathway. These findings underscore the potential for substantial future changes in the basin’s hydrological balance and water resource availability due to climate change.

Accordingly, the following section presents a detailed evaluation of the spatial distribution of future precipitation trends, temporal comparisons across projection periods, and station-specific precipitation changes throughout the basin.

3.4. Changes in Future Precipitation

Future precipitation changes in the Büyük Menderes Basin were evaluated using bias-corrected RF–based statistical downscaling under the SSP2-4.5 and SSP5-8.5 scenarios. Projections for 23 meteorological stations were analyzed across three future periods: near future (2025–2050), mid-century (2051–2075), and late century (2076–2099). The results indicate a consistent drying trend throughout the basin under both emissions scenarios.

Figure 5 presents the distribution of annual total precipitation from 1980 to 2014, the observational period, with values ranging from 210 mm to 1305 mm across the basin. Precipitation amounts are higher at western and southern stations, including Söke, Kale, Kavaklıdere, and Yatağan, while lower values are observed at inland and eastern stations such as Sarayköy, Sandıklı, Hocalar, Çivril, and Dinar. This spatial distribution reflects the influence of orographic precipitation, shaped by the basin’s topographic gradients and varying proximity to the Aegean Sea.

Figure 6 and Figure 7 display boxplots of future precipitation projections for 2025–2099 under the SSP2-4.5 and SSP5-8.5 scenarios. Compared to the observational period, SSP2-4.5 projections show a general decrease in precipitation. Precipitation remains relatively stable along the western coast, while inland areas such as Nazilli, Kuyucak, and Sarayköy experience more significant declines. These findings suggest that a basin-wide drying trend may develop even under moderate emissions scenarios.

Comparison of future projections with the observational period reveals a general decline in both median and mean annual precipitation throughout the 21st century. SSP2-4.5 results in a moderate drying signal across the basin, whereas SSP5-8.5 leads to substantial reductions in both the distribution range and mean precipitation at all stations. In the late-century period, precipitation decreases reach 15–21%, with the most pronounced declines observed in sub-basins around Nazilli, Sarayköy, Çal, and Güney.

Figure 8 and Figure 9 provide an illustrative station-level time-series view of annual precipitation projections for the Yatağan station. Equivalent annual projection plots were produced for all stations; however, only Yatağan is shown here to maintain a concise presentation. Under SSP2-4.5 in Figure 8, projected annual totals retain substantial interannual variability across the near (2025–2050), mid-century (2051–2075), and late-century (2076–2099) windows, while exhibiting a gradual downward shift from the near to the late period. This shift is reflected by a lower occurrence of years reaching the upper end of historical annual totals in the late-century window, consistent with a modest drying tendency under intermediate forcing.

Under SSP5-8.5 in Figure 9, the drying signal at Yatağan becomes stronger and more persistent, with the annual totals more frequently concentrated toward lower values and fewer late-century years approaching the higher annual amounts characteristic of the historical period. Compared with SSP2-4.5, the late-century window shows a clearer downward displacement of the annual distribution, indicating an intensification of drying under stronger radiative forcing. Together, these station-level trajectories complement the summary change statistics by showing how scenario-dependent differences manifest in the annual sequence while preserving interannual variability.

Overall, the Yatağan example illustrates that the projected changes are expressed not only as shifts in period-average precipitation but also through the year-to-year sequence, where interannual variability remains substantial while the central tendency moves toward lower annual totals with increasing forcing and time horizon. This time-series perspective provides additional context for interpreting scenario- and period-dependent differences, particularly regarding the persistence of drier conditions across successive years.

Figure 10 illustrates the spatial distribution of projected percentage changes in precipitation relative to the observational period. Furthermore, the numerical values of these percentage changes for each station are presented in Table 6 and Table 7 for the SSP245 and SSP585 scenarios, respectively. Under SSP2-4.5, marginal increases are projected for the western coastal areas in the near future, while a general drying trend is anticipated in inland and eastern regions. This trend intensifies in the mid- and late-century periods, with reductions of up to 10–15% in the central basin. These results indicate that, even under moderate emissions, precipitation is projected to decline progressively from the interior toward the east, with an increasing drying tendency.

Under the high-emission scenario, the drying signal is substantially more pronounced. Increased greenhouse gas concentrations result in reductions of approximately 10% in most parts of the basin, even in the near future. In SSP5-8.5 projections, these declines intensify to 15–22% during the mid- and late-century periods and extend across the entire basin. These results demonstrate the basin’s high sensitivity to climate change. Both scenarios project a systematic decrease in precipitation, with the drying tendency doubling under high emissions. This projected decline, significant in both magnitude and duration, is likely to have profound impacts on the basin’s future hydrological regime. The findings highlight considerable risks for sustainable water resource management, agricultural planning, and hydropower potential in the Büyük Menderes Basin under future climate change.

4. Discussion

4.1. Key Findings and Methodological Advantages

Precipitation projections were assessed for near-term (2025–2050), mid-century (2051–2075), and late-century (2076–2099) periods under SSP2-4.5 and SSP5-8.5 scenarios. The results indicate a consistent basin-wide decline in precipitation for both scenarios. SSP2-4.5 projects moderate reductions, whereas the high-emission SSP5-8.5 scenario forecasts decreases of up to 15–22%. These findings suggest substantial shifts in precipitation regimes are probable throughout much of the basin, particularly under SSP5-8.5. The most significant reductions are anticipated in sub-basins near Aydın, Nazilli, Sarayköy, Güney, and Çal.

Although the overall workflow is widely used in climate change impact studies, the present framework incorporates several features that improve physical consistency and station-scale applicability. First, predictors are selected in a station-specific manner (MPSP/APREG), rather than relying on a fixed set of large-scale variables across the basin, which supports robust local skill. Second, GCM fields are harmonized with ERA5 using the IDW–Delta–Variance approach prior to downscaling, reducing scale and distribution mismatches between the driving model and the reanalysis-based training domain. Third, the RF algorithm captures multivariate and nonlinear relationships relevant to monthly precipitation while allowing rigorous evaluation through OOB and independent testing. Finally, trend-preserving QDM is applied at the station scale to correct residual distributional biases without undermining the climate-change signal, which is a known limitation of conventional bias-correction approaches. From an applications perspective, reliable station-scale precipitation projections are also critical for reservoir operation and hydropower planning in the Büyük Menderes Basin; recent optimization and management studies for the basin emphasize that robust future inflow estimates are a key limiting factor [103,104].

4.2. Comparison with Previous Studies

Our basin-scale drying signal is largely consistent with recent CMIP6-based assessments over Türkiye that highlight increasing hydroclimatic stress, particularly across the Aegean–Mediterranean transition zone. For example, a national-scale analysis of downscaled CMIP6 projections reported increasingly dry conditions toward the late century and noted that total precipitation can decline by up to ~20% over the Aegean and Mediterranean regions under SSP5-8.5, implying stronger water-stress conditions compared with SSP2-4.5 [105]. Complementarily, aridity-focused CMIP6 analyses show that multiple aridity indices consistently indicate a shift toward drier climate regimes in Türkiye after the mid-century, with particularly notable increases projected for the inner Aegean among other regions—supporting the interpretation that precipitation reductions in western Anatolia are part of a broader, robust drying tendency [106]. At the same time, model evaluation studies emphasize that CMIP6 generally improves precipitation skill relative to CMIP5 while still exhibiting non-negligible model spread and regional differences, reinforcing that our results should be interpreted as conditional on the selected GCM and the applied downscaling/bias-adjustment chain [65]. Finally, impact-oriented studies from nearby southwestern basins such as the Burdur Basin that incorporate CMIP6 information similarly point to heightened future drought-related pressures on water systems, which is qualitatively aligned with the drying implications derived here for the Büyük Menderes Basin [107].

At the basin scale, our findings also agree with recent CMIP6-driven drought analyses specifically targeting the Büyük Menderes Basin. Using multiple CMIP6 GCMs under SSP2-4.5 and SSP5-8.5 and computing SPEI after bias correction, Rotbeei et al. projected an increase in dry-month occurrence toward the late century, supporting the likelihood of intensifying drought conditions under stronger forcing [55]. Observational analyses further contextualize this signal: trend-based evaluations over the Aegean region (including the Büyük Menderes) using multi-station records identify clear regional warming over recent decades, a key co-driver that can amplify drought risk even when precipitation changes are spatially heterogeneous [108]. Importantly, our mean/median precipitation decreases do not preclude intensification of short-duration rainfall extremes—consistent with national-scale CMIP6 downscaling work showing increased return levels of daily maximum precipitation, with substantially larger increases under SSP5-8.5 [6]. In addition, multi-model CMIP6 basin assessments employ ERA5 as a reference underscore the value of ensemble-based evaluation for quantifying inter-model uncertainty in precipitation and temperature projections, which complements our (single-model) methodological framework and motivates future extension to multi-model ensembles [60]. Finally, basin-scale hydroclimatic impact studies coupling bias-corrected climate projections with hydrological modeling in Türkiye frequently report seasonal redistribution (wetter wet seasons and drier dry seasons), emphasizing that water-management implications depend not only on annual totals but also on intra-annual shifts—an aspect that is consistent with the spatially heterogeneous changes identified in our downscaled station-scale projections [102]. Historical analyses for the Aegean region, including the Büyük Menderes Basin, also indicate pronounced drying [109].

4.3. Limitations and Future Work

A limitation arises from the coarse native resolution of HadGEM3-GC31-LL over the basin, where only a small number of GCM grid cells are available. This may smooth local gradients and under-represent orographic and coastal-transition effects, potentially affecting station-to-station contrasts and extremes. Because the basin is represented by only five native HadGEM3 grid cells, the regridded GCM predictors largely reflect an interpolated large-scale signal with limited independent spatial degrees of freedom. Therefore, fine-scale spatial contrasts among stations should be interpreted primarily as arising from the station-specific RF transfer function rather than from resolved GCM-scale gradients, and projected local differences, especially for extremes, remain more uncertain. In the proposed workflow, this scale mismatch is partly mitigated because the RF models are trained using higher-resolution ERA5 predictors together with station observations, and GCM fields are first harmonized to ERA5 statistics using the IDW–Delta–Variance approach prior to downscaling. Nevertheless, residual structural uncertainty related to coarse GCM sampling remains.

The observational dataset used in this study covers 1980–2014; post-2014 observations were not included due to data-availability constraints. Here, 2025 denotes the start of the scenario projection period rather than an observational endpoint.

In addition, projections were generated using a single CMIP6 realization (r1i1p1f3) rather than an ensemble mean, and therefore internal variability and ensemble spread are not quantified. Accordingly, the projected changes should be interpreted as conditional on the selected model configuration. In addition, future work will explore the sensitivity of projections to predictor choice and downscaling configurations, particularly for extremes. Extending the framework to multi-model and multi-member ensembles and higher-resolution simulations is a natural next step.

5. Conclusions

This study developed station-scale monthly precipitation projections for the Büyük Menderes Basin in western Türkiye using a hybrid statistical downscaling framework that combines Random Forest modeling with trend-preserving quantile delta mapping. ERA5 reanalysis predictors and station observations were used to train and evaluate the downscaling models, and CMIP6 HadGEM3-GC31-LL outputs were subsequently harmonized and downscaled to generate projections for SSP2-4.5 and SSP5-8.5. Model evaluation indicates overall good to very good performance across stations.

Across the basin, projections suggest a consistent tendency toward lower precipitation totals, with moderate reductions under SSP2-4.5 and larger decreases under SSP5-8.5 reaching approximately 15–22% by the late century. The strongest reductions are concentrated in several sub-basins, particularly around Aydın, Nazilli, Sarayköy, Güney, and Çal, indicating spatially heterogeneous but broadly coherent drying patterns. These results highlight increasing hydroclimatic stress and emphasize the need to consider climate-informed planning for water-dependent sectors in the basin. The framework is readily extensible to other semi-arid basins in Türkiye, subject to data availability and local validation.

Limitations include the use of a single driving GCM at coarse native resolution and the lack of explicit inter-model/ensemble uncertainty; therefore, the projections should be interpreted as conditional on the selected CMIP6 configuration. Future work will extend the same framework to multi-model and multi-member ensembles and higher-resolution simulations to quantify uncertainty and strengthen robustness.