Fuzzy Modeling Strategies for Groundwater Level Forecasting: Comparing Local, Integrated, and Behavioral Frameworks for a Data-Limited Coastal Aquifer in the Eastern Mediterranean

Abstract

Groundwater modeling in semi-arid regions presents significant challenges due to complex aquifer dynamics, limited data availability, and heterogeneous hydrogeological conditions. This study presents a comprehensive comparative analysis of three fuzzy expert system strategies for monthly groundwater level forecasting in the Al-Hsain Basin, Syria: localized models based on hydrogeographical grouping, a unified basin-wide approach, and an innovative behavioral clustering methodology. Using synchronized rainfall and temperature data from 35 monitoring wells over four years (2020–2024), we developed and evaluated fuzzy inference systems’ directional classification accuracy as the primary performance metric, categorizing groundwater level changes into rise, stable, and decline states rather than predicting continuous values. This choice reflects the qualitative nature of fuzzy expert systems and their suitability for groundwater management under data-limited conditions. The behavioral clustering approach achieved excellent overall performance with a mean accuracy of 0.74, outperforming localized models (0.71) and unified models (0.67). Behavioral clustering demonstrated effectiveness in 66% of wells, with individual accuracy improvements reaching up to 0.23, while reducing model complexity from five group-specific systems to three behaviorally coherent clusters. Localized models achieved optimal performance in 29% of wells where hydrogeological conditions aligned with spatial assumptions, whereas unified models provided consistent moderate performance across 89% of locations. The incorporation of lagged variables and seasonal indices in behavioral clustering models proved essential for capturing temporal complexity in semi-arid groundwater responses. Statistical analysis revealed lower intra-group variability in behavioral clusters (standard deviation 0.06–0.09) than in geographical groupings (0.08–0.14), confirming improved functional homogeneity through response-based organization. These findings indicate that fuzzy modeling strategy selection should be context-dependent, with behavioral clustering offering an effective balance between accuracy, interpretability, and generalization for regional groundwater management applications. The novelty of this work lies in isolating the effect of fuzzy system organization logic (localized, unified, and behavioral) on forecasting performance, robustness, and transferability, evaluated under an identical inference and time-series validation framework.

1. Introduction

Groundwater constitutes a fundamental component of global water resources, supplying nearly half of the world’s drinking water and sustaining agricultural production across diverse climatic regions [1,2,3]. Acting as natural subsurface reservoirs, aquifers provide resilience against surface water variability and seasonal scarcity. However, accelerating groundwater depletion driven by population growth, intensive irrigation, and climate change has raised serious concerns regarding long-term sustainability [4,5,6,7]. Recent global assessments indicate that groundwater systems are increasingly unable to buffer hydrological extremes, with declining storage trends observed in many major aquifers worldwide [2,4,5].

The challenge of groundwater management is particularly acute in semi-arid regions, where aquifers often serve as the primary or sole freshwater resource [8,9]. These environments are characterized by highly variable precipitation, elevated evapotranspiration, and episodic recharge processes, resulting in complex and delayed groundwater responses [10,11,12]. Observational and modeling studies demonstrate that groundwater recharge in semi-arid regions depends disproportionately on infrequent high-intensity rainfall events, leading to strong nonlinearity and temporal lag effects in groundwater level dynamics [8,9,12]. Climate change projections further suggest that rising temperatures and altered precipitation regimes will exacerbate groundwater stress in many semi-arid basins [1,10,13]. The Middle East exemplifies these challenges, with widespread groundwater depletion documented across regional aquifers and severe impacts observed during recent drought periods, such as the 2007–2010 Levantine drought [14,15]. Accurate groundwater level forecasting is therefore essential for sustainable water resource management in semi-arid environments. However, predictive modeling remains challenging due to hydrogeological heterogeneity, nonlinear climate–groundwater interactions, delayed system responses, and limited monitoring data [13,16,17]. These difficulties are particularly pronounced in developing or conflict-affected regions, where long-term datasets required for conventional modeling approaches are often unavailable.

The Al-Hsain Basin in coastal Syria represents a typical semi-arid Mediterranean aquifer system facing these challenges. The basin exhibits pronounced seasonal climatic variability and complex geological conditions, resulting in spatially and temporally heterogeneous groundwater responses [18,19]. Declining groundwater levels across the basin highlight the urgent need for reliable, data-efficient forecasting tools to support sustainable management under uncertainty. In response to these challenges, a growing body of research has explored alternative modeling approaches that can operate effectively under data-limited conditions. The following section reviews existing groundwater modeling methodologies, critically evaluates their limitations, and identifies the specific research gap addressed by the present study.

Accordingly, this study aims to systematically evaluate three fuzzy expert system strategies for groundwater level forecasting in a semi-arid, data-limited coastal aquifer: (i) localized hydrogeographical models, (ii) a unified basin-wide model, and (iii) a behavioral clustering-based approach. Specifically, the study addresses the following research objectives: (1) to assess the predictive performance of each modeling strategy using a consistent validation framework; (2) to examine the extent to which behavioral clustering improves accuracy and internal consistency compared to spatial grouping; and (3) to evaluate the trade-offs between model complexity, generalization capability, and practical applicability for groundwater management under data-limited conditions.

Beyond performance comparison, the methodological contribution of this study lies in the explicit evaluation of alternative fuzzy expert system organization logics under a unified inference and validation framework. While most existing fuzzy-based groundwater studies adopt a single modeling structure, this work systematically contrasts spatially localized, basin-wide unified, and behavior-driven clustering strategies using identical data, rule structures, and evaluation criteria. By isolating the effect of grouping logic on predictive performance, internal consistency, and model transferability, the proposed framework provides new methodological insight into the structure of fuzzy expert systems for heterogeneous, data-limited aquifer systems.

In practical terms, this contribution goes beyond applying fuzzy logic by providing a fully controlled and consistent comparison of alternative fuzzy organization strategies, introducing behavior-driven clustering for directional groundwater-state forecasting, and quantifying accuracy–complexity–transferability trade-offs under data-limited conditions.

2. Literature Review

Groundwater modeling has traditionally relied on physically based numerical models, such as MODFLOW, which simulate subsurface flow processes using governing equations and detailed representations of aquifer properties and boundary conditions [20]. These models provide valuable mechanistic insights into studies on groundwater dynamics and transport. However, their predictive reliability depends heavily on the availability of detailed hydrogeological data and well-constrained parameterization. In heterogeneous aquifer systems, geological uncertainty and unknown future boundary conditions can significantly compromise forecast accuracy [21,22,23]. To address these limitations, data-driven approaches have gained prominence in groundwater forecasting. Artificial neural networks (ANNs), support vector machines, and other machine learning techniques have demonstrated strong predictive performance by learning complex, nonlinear relationships directly from historical data [24,25,26]. Researchers have successfully applied these methods to groundwater-level forecasting across diverse hydroclimatic settings, including semi-arid regions [27,28]. The hybrid and ensemble approach further improves robustness and accuracy by combining multiple models or optimization frameworks [29,30,31].

Despite their predictive capabilities, many data-driven groundwater models operate as black boxes, offering limited interpretability and few opportunities to incorporate hydrogeological understanding or expert judgment [29,30]. This lack of transparency presents a significant limitation in water resource management, where decision-makers require models that are not only accurate but also explainable and trustworthy [31,32]. Both physically based and data-driven models often assume spatial homogeneity or apply uniform model structures across entire basins. Such assumptions frequently fail in semi-arid aquifers characterized by strong geological heterogeneity, variable recharge mechanisms, and localized anthropogenic influences. As a result, model performance can vary substantially between monitoring locations, reducing confidence in basin-wide predictions [32].

Fuzzy-logic–based models have emerged as a valuable alternative for groundwater applications, particularly under conditions of uncertainty and data scarcity [30]. By employing linguistic variables and rule-based reasoning, fuzzy expert systems explicitly represent uncertainty and integrate expert knowledge into the modeling process. Compared to neural networks, fuzzy systems offer superior interpretability through transparent “if–then” rules that practitioners can examine and adjust [33]. Numerous studies have demonstrated the effectiveness of fuzzy logic in groundwater-related applications, including groundwater potential mapping, recharge assessment, and level forecasting, especially in semi-arid and Mediterranean environments [34,35]. More recent research has explored hybrid fuzzy systems that combine fuzzy inference with clustering, evolutionary algorithms, or ensemble frameworks to enhance predictive performance and robustness [36,37,38,39]. These studies suggest that incorporating functional similarity and adaptive structures can improve model reliability in complex hydrological systems.

While modern data-driven and machine-learning approaches have demonstrated strong predictive capabilities in groundwater modeling, their reliance on data-intensive training procedures and black-box structures often limits interpretability and transferability, particularly in data-scarce environments. In contrast, fuzzy expert systems offer transparent rule-based reasoning, but their implementation relies on a single grouping logic, most commonly based on spatial proximity or predefined hydrogeological classes. As a result, existing studies rarely examine how alternative fuzzy grouping strategies influence predictive performance, internal consistency, and generalization behavior in heterogeneous aquifer systems. This lack of critical synthesis between fuzzy grouping strategies and modern data-driven approaches represents a key gap in the current literature.

While existing studies confirm the suitability of fuzzy logic for groundwater modeling, most applications adopt a single modeling strategy. Some studies employ localized fuzzy models tailored to specific hydrogeological zones, while others apply unified basin-wide frameworks that assume typical climate–groundwater relationships across all monitoring locations [31,36]. However, systematic comparisons between alternative fuzzy modeling strategies remain limited, particularly in semi-arid regions where groundwater responses exhibit pronounced spatial and temporal heterogeneity.

Moreover, although groundwater modelers have applied clustering techniques, their integration with fuzzy expert systems based explicitly on observed hydrological response behavior—rather than spatial proximity or geological classification—has received limited attention [37]. The relative effectiveness of localized, unified, and behavior-based fuzzy modeling strategies under consistent data and validation frameworks, therefore, remains insufficiently explored.

Addressing this gap is critical to developing practical, adaptable groundwater forecasting tools for semi-arid, data-limited environments. A systematic comparison of fuzzy modeling strategies can clarify the trade-offs between accuracy, generalization, and model complexity, providing actionable guidance for water managers operating under varying data and resource constraints [27,38]. By evaluating localized, unified, and behavioral clustering–based fuzzy expert systems using the same dataset and performance metrics, the present study advances methodological understanding of fuzzy groundwater modeling. It provides new insights into leveraging behavioral similarity to improve forecasting accuracy while maintaining interpretability and scalability. This approach is particularly relevant for heterogeneous aquifer systems such as the Al-Hsain Basin, where conventional spatial organization alone fails to capture dominant groundwater response mechanisms.

3. Materials and Methods

3.1. Study Area

The Al-Hsain Basin is located along the Mediterranean coastline of Tartous Governorate in western Syria and represents a typical semi-arid Mediterranean hydroclimatic environment. The basin extends eastward from the coastal plain into moderately elevated terrain and is predominantly agricultural, with irrigated crops and orchards interspersed with residential and light industrial activities. The regional climate exhibits distinct wet and dry seasons, with most of the annual precipitation concentrated between November and March. Summer months produce high temperatures and low rainfall, leading to strong evapotranspiration demand and prolonged periods of groundwater depletion.

Hydrogeologically, the basin comprises Quaternary alluvial deposits, Neogene sedimentary formations, and underlying Cretaceous bedrock [18,19]. This stratigraphic diversity creates a heterogeneous aquifer system, with unconfined conditions prevailing in shallow alluvial units and semi-confined to confined conditions in deeper formations. Groundwater represents the principal freshwater resource across the basin, particularly during extended dry periods when surface water availability is limited. Declining groundwater trends observed in several localities emphasize the need for reliable forecasting tools to support sustainable groundwater management. The spatial distribution of the 35 monitoring wells across five hydrogeological zones provides comprehensive coverage of the basin’s geological variability (Figure 1). Figure 2 further illustrates the geological and hydrogeological framework. The five hydrogeological zones referenced in Figure 2 reflect the dominant lithological and hydrostratigraphic units of the basin and are used only for descriptive purposes in this section.

3.2. Consideration of Seawater Intrusion

As the Al-Hsain Basin lies within a coastal setting, seawater intrusion may threaten water quality and require special consideration. Seawater intrusion can significantly affect groundwater levels in coastal aquifers, particularly under conditions of intensive pumping and prolonged groundwater decline.

This study did not explicitly model seawater intrusion because systematic salinity or electrical conductivity (EC) measurements were unavailable across the monitoring network. Nevertheless, several hydrogeological characteristics of the basin justify its exclusion as a dominant confounding factor. Most monitoring wells are located inland or at elevations exceeding mean sea level, where freshwater hydraulic heads are sufficient to prevent lateral seawater encroachment. In addition, the aquifer system receives substantial seasonal freshwater recharge during the wet period, maintaining a seaward hydraulic gradient that limits saline water intrusion and upward movement.

Furthermore, the observed groundwater level fluctuations exhibit strong seasonal coherence with precipitation and temperature patterns, indicating that climatic forcing rather than density-driven saline effects dominates groundwater dynamics at the monthly temporal scale considered in this study. Consequently, while seawater intrusion may locally affect groundwater quality near the coastline, it is unlikely to exert a significant influence on the monthly groundwater level changes (ΔGWL) analyzed herein.

3.3. Data Collection and Preprocessing

3.3.1. Meteorological and Groundwater Data

The empirical basis of this study comprises synchronized time series of climatic forcing variables and groundwater level observations collected over 4 years (January 2020–March 2024). Monthly precipitation totals and mean temperatures were obtained from the Syrian General Directorate of Meteorology and represent the primary climatic drivers of groundwater recharge and evapotranspiration in semi-arid environments [41]. These variables are essential for characterizing groundwater storage dynamics, in which precipitation timing and intensity strongly influence recharge. Meteorological data are assumed to represent basin-scale climatic conditions.

Groundwater level data were collected from 35 observation wells distributed across the basin to capture hydrogeological variability. Wells are completed in Quaternary alluvial deposits, Neogene sedimentary formations, and Cretaceous units, ensuring representation of diverse aquifer conditions. Monthly water level measurements were obtained using calibrated electro-optical devices with a precision of ±0.5 cm, supported by field verification procedures. Groundwater level data were initially recorded as depth-to-water (DTW) and subsequently converted to monthly level changes (ΔGWL) to represent the aquifer’s directional response.

Geographic coordinates, surface elevation, total depth, and lithological classification characterize each well. At this stage, wells are independent observation points without predefined categorization or grouping.

To provide a preliminary overview of the basin’s hydrologic behavior, Figure 3 illustrates the monthly groundwater level (GWL) time series for five representative wells (w22, w30, w15, w12, and w11), each selected from one of the five hydrogeographical groups. These time series demonstrate spatial diversity in aquifer responses, with varying degrees of seasonal fluctuation and response lags that correlate with the climatic forcing variables shown in Figure 3.

Climatic and groundwater datasets were temporally synchronized to a uniform monthly time step suitable for developing a fuzzy inference system. The synchronized climatic record (Figure 4) exhibits pronounced Mediterranean semi-arid seasonality, with approximately 75–80% of annual precipitation occurring between November and March and peak monthly totals of 150–200 mm. Summer months (June–September) are characterized by near-zero precipitation and elevated temperatures, resulting in episodic recharge followed by extended periods of depletion driven by high evapotranspiration. This strong seasonality and episodic recharge behavior introduce nonlinearity and lag effects in groundwater responses, complicating conventional modeling and motivating the use of fuzzy logic approaches capable of representing climate–aquifer uncertainty and nonlinearity [30,31]. These synchronized input–output time series provide the baseline dataset for subsequent preprocessing and fuzzy modeling.

3.3.2. Quality Control and Preprocessing

Before model development, the assembled datasets underwent systematic quality control and preprocessing to ensure consistency, accuracy, and suitability for subsequent fuzzy modeling [27,28]. Missing values in the meteorological time series occurred infrequently and were addressed using linear interpolation to preserve temporal continuity while avoiding the introduction of artificial trends. Groundwater level data required more extensive quality control, as several wells exhibited occasional measurement gaps and anomalous values, particularly during early monitoring phases. Outliers were identified using interquartile range analysis, supplemented by visual inspection of time series plots to distinguish measurement errors from legitimate hydrological extremes. Suspected digitization errors were corrected when supporting evidence existed, while uncorrectable data points were removed from analysis. Wells displaying excessive missing data or erratic behavior inconsistent with physical expectations were excluded to maintain dataset integrity.

Explicit preprocessing rules, applied uniformly over all datasets, ensured reproducibility. Preserving the temporal continuity required for lag-based predictors required linear interpolation in areas where isolated gaps occurred in groundwater records. Longer consecutive gaps required removal from the feature extraction and model evaluation steps. Interquartile range criteria helped flag outliers by identifying values outside Q1 − 1.5 × IQR or Q3 + 1.5 × IQR. Subsequent inspections against neighboring observations helped distinguish measurement errors from physically plausible extremes. The process removed only values identified as clear measurement or digitization errors and retained all hydrologically consistent extremes. The analysis also excluded wells exhibiting extensive missing records or persistent erratic behavior inconsistent with physical expectations. The three modeling strategies received identical preprocessing rules to ensure that performance differences reflect the modeling strategy rather than data handling.

All input variables were normalized to the range [0, 1], facilitating the consistent definition of fuzzy membership functions and the uniform handling of variables with different units and scales [31]. This normalization step ensures numerical stability and comparability across climatic and groundwater inputs while preserving relative variability in the original data. A monthly seasonal index (values 1–12) was introduced as an additional input variable to represent calendar-based seasonal effects that are not fully captured by temperature variations alone. Consecutive monthly groundwater level observations provided the level change (ΔGWL) to the primary model output variable.

3.4. General Fuzzy Modeling Framework

The groundwater forecasting methodology adopted in this study is based on a fuzzy expert system framework designed to operate under conditions of hydrogeological heterogeneity and limited data availability. Figure 5 illustrates the overall structure of the framework, presenting the complete sequence of steps used to transform raw climatic and groundwater observations into model predictions, then evaluate alternative modeling strategies with a consistent methodology.

The framework begins by preparing synchronized climatic and groundwater datasets at a uniform monthly time resolution. Following data collection, quality control, and normalization, the processed inputs are prepared for fuzzy inference by ensuring numerical consistency across variables with different units and scales. These preparatory steps establish a reliable empirical basis for subsequent fuzzy modeling while preserving the essential temporal variability of climatic forcing and groundwater responses. Climatic variables, including monthly precipitation and mean temperature, constitute the primary external forcing inputs controlling groundwater recharge and depletion processes in the semi-arid study area. In selected modeling strategies, additional inputs derived from groundwater observations, such as lagged groundwater level changes and seasonal indices, are incorporated to capture temporal persistence and seasonal modulation of aquifer responses. These inputs are subsequently transformed into fuzzy linguistic variables, allowing gradual transitions between qualitative states and explicit representation of uncertainty inherent in climate–groundwater interactions. Expert-defined fuzzy rule bases are then applied to represent qualitative relationships between input conditions and groundwater level change responses. The fuzzy inference process evaluates the activated rules and aggregates their outputs to generate fuzzy representations of potential groundwater level change states. Defuzzification then converts aggregated fuzzy outputs into crisp estimates of monthly groundwater level change (ΔGWL), enabling direct comparison with observed data during model evaluation.

The pronounced hydrogeological heterogeneity of the Al-Hsain Basin, combined with the limited availability of long-term hydrogeological data, makes it particularly suitable for evaluating alternative fuzzy modeling strategies under realistic semi-arid conditions. Accordingly, the framework supports the implementation of multiple modeling strategies while maintaining a common inference structure and evaluation procedure.

All fuzzy modeling strategies implemented in this study share this identical general framework. Their strategies arise from preprocessing decisions on well grouping and the complexity of input variable sets used to generate fuzzy rule bases. By maintaining a consistent modeling approach across all strategies, the framework enables an objective comparison of predictive performance. It determines the influence of grouping logic and input selection on groundwater-level forecasting accuracy.

3.4.1. Fuzzy System Architecture

The fuzzy inference system implemented in this study follows a standard Mamdani-type architecture consisting of five sequential steps: (i) fuzzification of crisp climatic and groundwater inputs using triangular membership functions; (ii) evaluation of expert-defined IF–THEN rules linking input linguistic variables to groundwater response states; (iii) aggregation of activated rule outputs using the max–min inference mechanism; (iv) construction of aggregated fuzzy output sets representing possible groundwater level change states; and (v) defuzzification using the centroid method to obtain a crisp estimate of monthly groundwater level change (ΔGWL). This unified architecture is applied consistently across all three modeling strategies, ensuring that performance differences arise solely from grouping logic and input structure rather than inference mechanics.

A Mamdani-type fuzzy inference system provides the fuzzy modeling framework implemented in this study. Its advantage lies in its interpretability and suitability for expert-driven groundwater modeling under uncertainty. The architecture follows a structured sequence of fuzzification, rule evaluation, aggregation, and defuzzification, enabling transparent representation of nonlinear climate–groundwater relationships while accommodating imprecision inherent in hydrogeological processes.

Input variables, including monthly precipitation and mean temperature, constitute the primary climatic forcing factors driving groundwater recharge and depletion in the study area. In selected modeling strategies, additional inputs derived from groundwater observations—specifically lagged groundwater level change and a monthly seasonal index—are incorporated to capture temporal persistence and seasonal modulation of aquifer responses. The system output variable is the monthly groundwater level change (ΔGWL), which represents the aquifer’s directional response to climatic forcing.

During the fuzzification stage, predefined membership functions transform crisp input variables into fuzzy linguistic variables. This process allows individual observations to possess partial membership across multiple linguistic states, enabling smooth transitions between categories and avoiding sharp threshold effects. Such representation is particularly appropriate for semi-arid groundwater systems, where recharge and depletion processes exhibit gradual responses rather than abrupt changes.

The inference engine evaluates a set of expert-defined IF–THEN rules that link combinations of input linguistic states to corresponding groundwater response categories. Activated rules are aggregated to form fuzzy output sets representing potential groundwater level change states. Defuzzification is subsequently applied using the centroid method to convert aggregated fuzzy outputs into crisp ΔGWL estimates suitable for quantitative evaluation and comparison with observed data.

The Mamdani architecture ensures that each stage of the inference process—from input transformation to final prediction—remains transparent and interpretable. This feature distinguishes the proposed framework from black-box data-driven models and facilitates integration of expert knowledge, systematic comparison of modeling strategies, and practical application in groundwater management contexts.

3.4.2. Membership Functions and Fuzzification

Fuzzification is the process by which crisp input variables transform into fuzzy linguistic variables. In this study, triangular membership functions represented all input and output variables due to their simplicity, computational efficiency, and suitability for data-limited conditions. The membership functions were defined based on empirical percentiles of the observed data and expert knowledge of hydroclimatic conditions in semi-arid environments. Three linguistic categories divided rainfall and temperature inputs: Low, Medium, and High. Similarly, the groundwater level change (ΔGWL) output variable could Decline, remain Stable, or Rise. The parameters of the triangular membership functions, including lower bounds, peak values, and upper bounds, are summarized in

Table 1. Triangular membership function parameters for fuzzy linguistic variables.

Input Variable	Linguistic Term	Type	Lower Bound	Peak	Upper Bound
Rainfall	Low	Triangular	0	10	20
	Medium		10	25	40
	High		30	50	70
Temperature	Low		0	10	20
	Medium		10	20	30
	High		20	30	40
ΔGWL	Decline		−0.3	−0.2	−0.1
	Stable		−0.15	0	0.15
	Rise		0.1	0.2	0.3

, while Figure 6 shows their graphical representations.

The membership functions were designed with overlapping ranges to allow gradual transitions between linguistic categories and to avoid abrupt threshold effects. The degree of overlap—approximately 50% for rainfall and temperature, and 75% for ΔGWL—was selected empirically to balance sensitivity to input variations against excessive rule activation that could dilute inference precision [30]. This overlap ensures that input values close to category boundaries activate multiple linguistic terms simultaneously, rather than being assigned to a single class. As a result, multiple fuzzy rules may partially activate for a given input condition, which improves the system’s ability to represent uncertainty and smooth nonlinear relationships between climatic forcing and groundwater response.

Overlapping membership functions are particularly important in semi-arid groundwater systems, where recharge and depletion processes do not change abruptly but respond gradually to rainfall and temperature variations. By allowing partial membership across adjacent categories, the fuzzification process reduces sensitivity to small input variations and measurement uncertainty. This design improves the numerical stability of the inference process and enhances the robustness of model predictions.

Balancing interpretability, computational simplicity, and robustness under data-limited conditions guided the selection of triangular membership functions and the parameter ranges adopted. No alternative fuzzy partitions demonstrated consistent advantages over the selected setup.

All membership functions and defuzzification procedures are identical across the three modeling strategies to ensure methodological consistency and fair comparison.

3.4.3. Three Modeling Strategies

Three fuzzy modeling strategies demonstrated how different assumptions regarding groundwater system organization influence forecasting performance: localized hydrogeographical models, a unified basin-wide model, and behavioral clustering-based models. All strategies employ the same Mamdani-type fuzzy inference system architecture, membership functions, and defuzzification procedure described in the previous sections. The strategies differ in (i) the grouping method for monitoring wells, and (ii) the structure and complexity of the fuzzy rule bases used to represent climate–groundwater relationships.

Hydrogeographical grouping was defined based on a combination of spatial proximity to the main river, elevation, dominant geological formation, and inferred recharge conditions. Well groups reflected shared hydrogeological settings, including river-adjacent alluvial zones, inland Quaternary deposits, and older consolidated formations. Alternative classification schemes based solely on spatial distance or lithological units did not provide additional explanatory value beyond the adopted hydrogeographical grouping.

• Strategy I: Localized Fuzzy Models

In the localized modeling strategy, monitoring wells were grouped based on hydrogeological characteristics, including spatial proximity and dominant geological formations. Based on the geological framework of the Al-Hsain Basin and the spatial distribution of wells (Figure 1 and Figure 2), the 35 observation wells were divided into five hydrogeographical groups. For each hydrogeographical group, a separate fuzzy inference system was developed. A representative well was selected from each group based on data completeness and temporal stability. Fuzzy rule bases were derived from time series of representative wells and from expert interpretation of the relationships among rainfall, temperature, and observed groundwater level changes.

The localized rule bases employ precipitation and temperature as input variables and groundwater level change (ΔGWL) as the output variable. Rules are formulated using standard IF–THEN structures, such as “IF rainfall is Low AND temperature is High THEN ΔGWL is Decline,” reflecting physically reasonable groundwater responses under semi-arid conditions. The complete set of rules translates different combinations of climatic conditions into groundwater response categories.

Table 2. Localized fuzzy rule base for Group A wells.

Rule Number #	Rainfall	Temperature	ΔGWL
R1	Low	High	Decline
R2	Low	Medium	Decline
R3	Low	Low	Stable
R4	Medium	High	Decline
R5	Medium	Medium	Stable
R6	Medium	Low	Rise
R7	High	High	Stable
R8	High	Medium	Rise
R9	High	Low	Rise

presents an example of a localized fuzzy rule base.

Once defined, the group-specific rule bases were applied uniformly to all wells within the same hydrogeographical group without further modification. This approach assumes that a shared rule structure can represent wells within a group that exhibits similar responses to climatic forcing. The localized strategy allows flexibility in capturing spatial variations in aquifer properties and recharge mechanisms and may achieve higher accuracy at the local scale, although at the expense of reduced transferability across the basin [35].

• Strategy II: Unified Fuzzy Model

The unified modeling strategy develops a single fuzzy inference system for basin-wide application across all monitoring wells, regardless of spatial or geological differences [38]. This approach combines observations from all wells into a single modeling framework to identify generalized climate–groundwater relationships applicable across diverse hydrogeological conditions. A representative well exhibiting stable behavior, strong correlation with basin-wide trends, and complete data records provided the basis for rule extraction. The unified fuzzy rule base employs the same input variables as the localized models—monthly precipitation and temperature—but derives rules from patterns observed across the entire monitoring network rather than from individual hydrogeographical groups. The resulting rule structure, summarized in

Table 3. Unified fuzzy rule base for basin-wide application.

Rule Number #	Rainfall	Temperature	ΔGWL
R1	Low	High	Decline
R2	Low	Medium	Decline
R3	Low	Low	Stable
R4	Medium	High	Decline
R5	Medium	Medium	Stable
R6	Medium	Low	Rise
R7	High	High	Stable
R8	High	Medium	Rise
R9	High	Low	Rise

, represents fundamental climate–groundwater relationships assumed to operate consistently at the basin scale. The similarity between the unified and localized rule structures (Table 2) indicates that core climate–groundwater interactions remain broadly consistent despite geological heterogeneity. However, differences in rule activation may arise from local conditions and parameterization of the membership functions.

The unified rule base is applied uniformly to all monitoring wells to evaluate whether a simplified, generalized fuzzy system can achieve acceptable prediction accuracy in a heterogeneous semi-arid aquifer system. This approach offers practical advantages, including simplified implementation, reduced data requirements per location, and straightforward transferability to ungauged sites within the basin [31]. However, the assumption of basin-wide homogeneity may reduce prediction accuracy for wells exhibiting atypical behavior or influenced by localized geological or anthropogenic factors [21]. Evaluation of the unified model, therefore, provides insight into the trade-offs between generalization and accuracy in regional groundwater forecasting.

• Strategy III: Behavioral Clustering-Based Models

The behavioral clustering strategy groups monitoring wells based on observed hydrological response patterns rather than on spatial proximity or geological characteristics [37]. This approach presumes that shared fuzzy rule systems effectively model wells with similar temporal groundwater dynamics, even when geographically separated. By organizing wells according to functional similarity, the strategy aims to balance the specificity of localized models with the efficiency and scalability of unified approaches [42]. Implementation begins with extracting descriptive hydrological features from each well’s monthly groundwater level change (ΔGWL) time series. These features include statistical measures such as mean change, standard deviation, and temporal autocorrelation, as well as indicators of climatic sensitivity derived from correlations with rainfall and temperature variables [43,44,45]. Additional behavioral characteristics, including the frequency and persistence of rising or declining groundwater trends, were also considered. The resulting multidimensional feature sets served as input to an unsupervised K-means clustering algorithm, which partitioned the 35 monitoring wells into behaviorally coherent clusters. Three clusters were defined to ensure representation of dominant groundwater response behaviors while maintaining sufficient wells within each cluster for robust rule transfer and validation.

Objective cluster validation metrics provided the optimal number of behavioral clusters. They included the Elbow Method (within-cluster sum of squares), the Silhouette Coefficient, and the Davies–Bouldin Index, supported by principal component analysis (PCA) for visualization. In addition, sensitivity analysis revealed the influence of different lag orders (1–3 months) and alternative seasonal representations (calendar-based index, SPI-1, and SPI-3). Section 4.1 presents the results of the clustering validation and sensitivity analysis.

For each identified cluster, a representative well exhibiting typical response characteristics was selected to develop a cluster-specific fuzzy rule base. Unlike the localized and unified strategies, which rely solely on rainfall and temperature inputs, the behavioral clustering models incorporate an expanded input structure that includes lagged ΔGWL values and a monthly seasonal index. These additional variables allow the fuzzy system to account explicitly for temporal persistence and seasonal effects that temperature alone may not fully capture [12].

The behavioral rule bases encode complex temporal groundwater dynamics by linking current climatic conditions, antecedent groundwater states, and seasonal context to responses in groundwater levels. This structure reflects the memory effects inherent in groundwater systems, where current responses depend on both recent forcing and prior aquifer conditions [10].

Table 4. Behavioral clustering-based fuzzy rule base for Cluster 2.

Rule Number	Rainfall	Temp.	Lagged ΔGWL	Season	ΔGWL
R1	Low	High	Decline	Summer	Decline
R2	Low	Medium	Decline	Spring	Stable
R3	Medium	Medium	Stable	Autumn	Rise
R4	Medium	Low	Stable	Winter	Rise
R5	High	High	Rise	Summer	Stable
R6	High	Low	Decline	Winter	Rise
R7	Medium	High	Rise	Autumn	Stable
R8	Low	Low	Stable	Spring	Stable

presents an example of a cluster-specific fuzzy rule base developed for Cluster 2. The inclusion of seasonal indicators accounts for variations in recharge efficiency and evapotranspiration that may differ between seasons with similar temperatures, such as spring and autumn [45].

A two-stage workflow allows coupling between behavioral clustering and fuzzy modeling. The first stage groups monitoring wells into behaviorally coherent clusters based on similarity in groundwater response features derived from observed ΔGWL time series. The second stage selects a representative well within each behavioral cluster that exhibits typical response characteristics to construct a cluster-specific fuzzy rule base. This rule base is then applied uniformly to all remaining wells within the same cluster, without further tuning, to evaluate intra-cluster transferability and predictive consistency. In this way, clustering defines the functional grouping of wells, while fuzzy inference provides the predictive mechanism within each behavioral group.

The cluster-specific rule bases are transferred to all wells within the corresponding behavioral cluster without site-specific calibration. Once defined, the cluster-specific rule bases were applied uniformly to all remaining wells within the corresponding behavioral cluster to evaluate intra-cluster prediction performance. This strategy provides an intermediate modeling approach that offers enhanced specificity through behavior-based grouping while maintaining reasonable generalization and transferability across functionally similar wells [39]. It is particularly effective in settings where spatial or geological classifications fail to capture dominant groundwater controls, which may instead arise from complex interactions between climatic forcing, aquifer heterogeneity, and anthropogenic influences [27].

Table 5. Comparative overview of three fuzzy modeling strategies.

Aspect	Localized Strategy	Unified Strategy	Behavioral Clustering Strategy
Grouping principle	Spatial and geological similarity	Basin-wide (no grouping)	Functional similarity (responses)
Assumed similarity	Physical proximity	Universal climate–GWL relations	Empirical response behavior
Input variables	Rainfall, temperature	Rainfall, temperature	Rainfall, temperature, lagged ΔGWL, seasonal index
Rule base	Group-specific	Single basin-wide	Cluster-specific
Model complexity	Moderate	Low	High
Strength	Local precision	Simplicity and transferability	Captures temporal memory
Limitation	Reduced transferability	Reduced local accuracy	Higher data and complexity
Intended use	Physically coherent subregions	Rapid basin-scale assessment	Heterogeneous response systems

provides a systematic comparison of all three strategies, emphasizing their contrasting philosophies regarding spatial versus functional organization, input complexity, and the fundamental trade-off between local precision and general applicability. The localized approach assumes that physical proximity and geological similarity indicate hydrological similarity. The unified model assumes universal climate-groundwater relationships, while behavioral clustering empirically identifies functional similarities through observed response patterns, regardless of spatial or geological context [30,38].

3.4.4. Model Implementation

For each modeling strategy, the corresponding fuzzy rule base was applied to generate monthly groundwater level change predictions for wells not used in rule construction. Fuzzy inference was performed at each monthly time step using the defined membership functions and rule bases. All model implementations followed an identical computational workflow to ensure consistency across strategies and to isolate the effect of grouping logic and rule structure on model performance. At each monthly time step, crisp input variables were fuzzified, fuzzy rules were evaluated and aggregated, and centroid defuzzification was applied to produce ΔGWL predictions.

3.4.5. Model Validation

The predictive performance of each fuzzy modeling strategy was evaluated by systematically comparing model predictions with observed groundwater level changes across the monitoring network. Model validation employed classification accuracy as the primary performance metric, defined as the proportion of monthly predictions that correctly identified the direction of groundwater level change (rise, stable, or decline) relative to observations [20,23]. This categorical evaluation is consistent with the fuzzy system architecture, which produces qualitative classifications rather than precise numerical estimates, and provides an intuitive metric suitable for groundwater management applications.

To ensure independent validation and a realistic assessment of generalization capability, model testing was conducted using wells not involved in rule-based construction. For the localized strategy, rule bases derived from representative wells were applied to remaining wells within the same hydrogeographical group, thereby evaluating the assumption of hydrological similarity within spatial classifications. The unified strategy applied a single basin-wide rule base uniformly to all wells, testing the robustness of generalized climate–groundwater relationships under heterogeneous conditions [22,46]. Behavioral clustering models were validated through intra-cluster application, assessing whether wells grouped by similar response behavior exhibit consistent performance when modeled using cluster-specific fuzzy rules. Classification accuracy was computed as the percentage of correct directional predictions relative to total predictions, following the standard equation:

$$Accuracy (\%) =\left({\displaystyle \frac{Number of Correct Predictions }{Total Number of Predictions}}\right)\times 100$$

(1)

Unlike conventional machine learning approaches, fuzzy expert systems do not require explicit training or testing data splits, as they are rule-based and expert-driven [27]. Instead, model robustness was assessed by evaluating rule transferability across multiple independent wells, providing a realistic evaluation of operational performance in data-limited environments where site-specific calibration may not be feasible [28]. The comparative validation framework enables systematic assessment of the relative strengths and limitations of localized, unified, and behavioral clustering strategies. By applying consistent validation protocols and performance metrics, the analysis identifies trade-offs between model specificity and generalizability. It supports informed selection of appropriate fuzzy modeling strategies for groundwater level forecasting in semi-arid regions [17,34]. Based on the consistent implementation and validation framework described above, the three fuzzy modeling strategies can be directly compared in terms of predictive performance. The following section presents a comparative analysis of localized, unified, and behavioral clustering approaches.

3.5. Implementation Details and Software

Python (version 3.12) provided the coding base for the behavioral clustering and fuzzy inference systems. The K-means algorithm from the scikit-learn library generated functional grouping of monitoring wells, with cluster validity assessed using the Elbow Method, Silhouette Coefficient, and Davies–Bouldin Index. Fuzzy expert systems followed a Mamdani-type inference architecture. The architecture implemented membership function parameterization and centroid-based defuzzification through custom Python logic, supported by NumPy and SciPy for numerical computation.

The pandas and scikit-learn libraries performed data preprocessing, feature extraction, and evaluation—including balanced accuracy and Cohen’s Kappa. Custom Python scripts performed advanced validation procedures, including rolling-origin time-series evaluation and moving-block bootstrap resampling for uncertainty quantification. All analyses were executed on a standard desktop computing environment, ensuring that the proposed methodology is computationally efficient, reproducible, and does not require specialized hardware.

4. Results

4.1. Behavioral Clustering Validation and Sensitivity Analysis

To justify the selection of the optimal number of behavioral clusters, multiple cluster validation criteria were jointly employed, including the Elbow Method, the Silhouette Coefficient, and the Davies–Bouldin Index (DBI). As illustrated in Figure 7, the Elbow Method reveals a pronounced reduction in within-cluster sum of squares (WCSS) when increasing the number of clusters from K = 2 to K = 3, followed by a markedly diminished rate of improvement for higher K values. This pattern indicates that K = 3 represents a meaningful balance between cluster compactness and model parsimony.

Complementary validation using the Davies–Bouldin Index (Figure 8) further supports this choice, as DBI values decrease substantially up to K = 3. In contrast, subsequent reductions for higher K values are marginal and less stable. Although the Silhouette Coefficient exhibits a slight increase at higher cluster numbers (e.g., K = 5), the improvement remains limited and does not justify the additional model complexity introduced by increasing the number of behavioral groups.

Moreover, a principal component analysis (PCA) projection of the behavioral feature space (Figure 9) shows that three clusters provide a clear, interpretable separation of wells based on their hydrological response characteristics, particularly distinguishing wells with delayed and attenuated groundwater level responses from those with more rapid dynamics. From a modeling perspective, selecting K = 3 ensures a parsimonious yet physically interpretable clustering structure, minimizing the risk of overfitting and avoiding unnecessary proliferation of fuzzy rule bases in subsequent modeling stages. Accordingly, K = 3 was adopted as the optimal number of behavioral clusters for this study.

A sensitivity analysis evaluated the robustness of the behavioral clustering framework with respect to both the lag order applied to groundwater level changes (ΔGWL) and the definition of the seasonal index. The lag order directly reflects the groundwater system’s memory and delayed recharge processes, while the seasonal index captures the recurrent hydroclimatic forcing that governs groundwater fluctuations.

Physically, the lagged groundwater level change (ΔGWL_{t − 1}) represents aquifer memory, reflecting the fact that groundwater systems respond to climatic forcing with delayed and attenuated signals due to unsaturated-zone percolation, storage effects, and the time required for recharge fronts to propagate to the water table. In this context, a one-month lag captures the dominant short-term persistence in monthly observations without introducing unnecessary complexity. The seasonal index represents recurring intra-annual controls on recharge efficiency and discharge, including evapotranspiration demand, soil moisture replenishment, and seasonal pumping practices, which temperature alone may not fully represent. Together, these predictors allow the fuzzy rule base to encode delayed recharge processes and seasonal modulation of groundwater dynamics in a data-limited setting.

First, the sensitivity of clustering performance to the lag order was examined by testing lag values of 1, 2, and 3 months using a calendar-based seasonal representation. As shown in Figure 10, the clustering quality remains relatively stable across the tested lag orders, indicating that the behavioral grouping is not overly sensitive to the specific lag choice. While the Davies–Bouldin index shows a marginal improvement at a three-month lag, the silhouette coefficient attains its maximum at a one-month lag, suggesting a better balance between cluster compactness and separation. From a physical perspective, a one-month lag is also more consistent with the expected response time of shallow coastal aquifers, where recharge signals typically propagate to groundwater levels within short temporal scales. Accordingly, a one-month lag provided a parsimonious and physically meaningful choice.

Second, the relevance of the calendar-based seasonal index was assessed by comparing it with climate-based seasonality indicators, namely the Standardized Precipitation Index at the 1- and 3-month aggregation scales (SPI-1 and SPI-3). Figure 11 demonstrates that the calendar-based seasonal representation consistently yields higher silhouette values and lower Davies–Bouldin indices compared to both SPI-based alternatives. The results indicate that explicit calendar seasonality provides a more effective descriptor of groundwater response patterns in the study area. Physically, groundwater dynamics in the basin are governed not only by precipitation anomalies but also by seasonally recurring processes such as evapotranspiration, soil moisture depletion, and delayed recharge, which a calendar-based seasonal index captures directly.

Overall, the sensitivity analysis confirms that the adopted configuration—using a one-month lag for ΔGWL and a calendar-based seasonal index—offers the most robust, interpretable, and physically consistent representation of groundwater behavioral variability for subsequent fuzzy modeling.

4.2. Individual Strategy Performance

The predictive performance of the fuzzy modeling strategies was evaluated using classification accuracy, defined as the proportion of correctly predicted monthly groundwater level change directions (rise, stable, or decline). The same validation framework assessed the localized, unified, and behavioral clustering strategies using k to ensure consistent comparison. All reported accuracy values were calculated based on independent wells that were not used in rule-based construction, as described in Section 3.4.5. The results presented in this section provide a basis for comparing the effectiveness of the three modeling strategies and assessing their suitability for groundwater-level forecasting in semi-arid environments.

4.2.1. Localized Models

The localized fuzzy modeling strategy showed variable predictive performance across the five hydrogeographical groups. Classification accuracy at the well level ranged from 0.61 to 0.80 (

Table 6. Classification accuracy of local fuzzy models across 35 monitoring wells by hydrogeographical group. Accuracy values represent classification accuracy.

Group A		Group B		Group C		Group D		Group E
Well	Accuracy	Well	Accuracy	Well	Accuracy	Well	Accuracy	Well	Accuracy
w1	0.73	w4	0.71	w3	0.62	w24	0.77	w14	0.77
w2	0.75	w5	0.64	w11	0.77	w27	0.72	w15	0.76
w6	0.69	w13	0.72	w12	0.80	w28	0.71	w16	0.79
w7	0.79	w25	0.73	w19	0.66	w29	0.76	w17	0.72
w8	0.75	w30	0.77	w20	0.72	w31	0.66	w18	0.76
w9	0.68			w23	0.62	w32	0.64
w10	0.65			w26	0.68	w34	0.68
w21	0.65			w33	0.61	w35	0.75
w22	0.73

), with a basin-wide mean accuracy of 0.71. These results indicate that localized models are generally effective in representing climate–groundwater relationships within geographically coherent zones. However, noticeable differences in performance were observed both between groups and among wells within the same group, reflecting the inherent heterogeneity of the aquifer system.

Group-level average accuracies ranged from 0.68 to 0.76 (Figure 12). Group E achieved the highest average accuracy (0.76), with most wells exceeding 0.72, while Group C showed the widest performance range, with accuracies ranging from 0.61 to 0.80. The relatively strong performance of Group E is consistent with the presence of more stable Cretaceous formations and lower anthropogenic influence, which result in more predictable groundwater responses to climatic forcing.

The spatial distribution of well-level accuracy (Figure 13) reveals patterns that hydrogeographical grouping alone does not fully explain. High-performing wells (accuracy > 0.75) occur in several groups and over most geological zones. The highest accuracies were observed for wells w12 (0.80), w7 (0.79), w16 (0.79), w30 (0.78), w15 (0.76), and w18 (0.76), distributed across Groups A, C, D, and E. In contrast, lower-performing wells (accuracy < 0.65), including w3, w10, w21, w20, w23, w32, and w33, are scattered across Groups A, C, and D. These results indicate that hydrogeographical grouping captures some important controls on groundwater behavior but does not fully account for local variability caused by complex geological conditions or anthropogenic influences. While localized models provide good performance in many areas, their effectiveness varies depending on site-specific response characteristics.

4.2.2. Unified Models

The unified fuzzy modeling strategy produced more uniform predictive performance across the basin compared to the localized approach. Well-level classification accuracy ranged from 0.58 to 0.76, with a mean accuracy of 0.67 (

Table 7. Classification accuracy of the unified fuzzy model organized by hydrogeographical groups. Accuracy values represent classification accuracy.

Group A		Group B		Group C		Group D		Group E
Well	Accuracy	Well	Accuracy	Well	Accuracy	Well	Accuracy	Well	Accuracy
w1	0.68	w4	0.64	w3	0.61	w26	0.72	w14	0.58
w2	0.60	w5	0.75	w11	0.66	w27	0.68	w15	0.75
w6	0.66	w13	0.59	w12	0.66	w28	0.64	w16	0.63
w7	0.73	w24	0.71	w19	0.73	w34	0.66	w17	0.68
w8	0.68	w25	0.63	w20	0.74	w35	0.74	w18	0.67
w9	0.67	w29	0.68	w23	0.73			w31	0.61
w10	0.62	w30	0.74	w32	0.66
w21	0.64			w33	0.62
w22	0.76

). Although this represents a 4% reduction relative to the localized models, the unified strategy showed reduced variability across hydrogeographical groups. Group-average accuracies differed by only 0.07 between the highest-performing group (Group A, 0.68) and the lowest-performing group (Group E, 0.65) (Figure 14). This limited variation reflects the use of a single basin-wide fuzzy rule base without local calibration, yielding consistent but moderate performance across diverse hydrogeological settings.

Several wells achieved relatively high accuracy under the unified strategy, including w22 (0.76), w5 (0.75), w15 (0.75), w20 (0.74), w30 (0.74), and w35 (0.74). These wells are distributed across multiple hydrogeographical groups, indicating that their groundwater responses align well with the generalized climate–groundwater relationships captured by the unified rule base. In contrast, lower accuracies were observed at wells such as w14 (0.58), w13 (0.59), and w2 (0.60), suggesting the influence of local geological or anthropogenic factors not represented in the basin-wide model.

The spatial distribution of accuracy for the unified model (Figure 15) is notably uniform, with fewer localized clusters of very high or very low performance compared to the localized strategy. This pattern indicates that the unified model effectively generalizes climate–groundwater relationships across the basin, albeit with reduced sensitivity to site-specific variability. As a result, the unified approach is well-suited for basin-scale assessments where simplified implementation and broad applicability outweigh local optimization.

4.2.3. Behavioral Clustering

The behavioral clustering strategy produced the broadest range of well-level performance among the three approaches, with classification accuracies ranging from 0.60 to 0.85. Despite this variability, it achieved the highest overall mean accuracy of 0.74 (

Table 8. Classification accuracy of behavioral clustering models by cluster assignment. Accuracy values represent classification accuracy.

Cluster 1		Cluster 2		Cluster 3
Well	Accuracy	Well	Accuracy	Well	Accuracy
w2	0.70	w8	0.60	w1	0.75
w5	0.81	w15	0.77	w3	0.85
w6	0.68	w17	0.81	w4	0.72
w7	0.71	w21	0.79	w9	0.84
w13	0.66	w23	0.65	w10	0.74
w16	0.66	w24	0.79	w11	0.70
w18	0.64	w26	0.81	w12	0.77
w20	0.65	w27	0.74	w14	0.84
w22	0.71	w33	0.73	w19	0.82
w25	0.61	w34	0.83	w28	0.80
				w29	0.73
				w30	0.78
				w31	0.68
				w32	0.74
				w35	0.81

). The K-means clustering procedure partitioned the 35 monitoring wells into three behavioral clusters based on similarities in groundwater response characteristics, independent of spatial location or geological classification.

Cluster-level mean accuracies show clear performance differences among the three groups (Figure 16). Cluster 3 achieved the highest mean accuracy (0.77), followed by Cluster 2 (0.75) and Cluster 1 (0.68). Cluster 3 includes 15 wells distributed across several hydrogeographical groups and shows strong internal consistency, with 73% achieving accuracies above 0.75. These wells exhibit relatively stable and climatically responsive groundwater behavior, characterized by consistent seasonal patterns and moderate lag effects.

Cluster 2 consists of 10 wells and shows high, consistent performance, with 80% exceeding an accuracy of 0.70. The expanded fuzzy rule structure used for this cluster, which incorporates lagged groundwater level change (ΔGWL) and a seasonal index, appears effective in capturing wells with more complex temporal dynamics. The highest-performing wells in this cluster include w34 (0.83), w26 (0.81), and w17 (0.81), all of which exhibit strong memory effects and clear seasonal responses.

Cluster 1 also contains 10 wells but exhibits the greatest performance variability, with accuracies ranging from 0.61 to 0.81. This cluster appears to include wells with irregular or transitional groundwater behavior, potentially influenced by multiple hydrogeological processes or localized anthropogenic effects that the climatic input variables may not fully represent.

The spatial distribution of behavioral clusters (Figure 17) demonstrates that wells with similar groundwater response behavior are distributed across multiple hydrogeographical groups. This pattern confirms that neither spatial proximity nor geological boundaries constrain functional similarity. High-performing wells from all three clusters are spread throughout the basin, indicating that behavioral classification provides a useful framework for capturing groundwater response characteristics that spatial or geological grouping alone may not adequately describe.

4.3. Comparative Analysis

4.3.1. Cross-Strategy Comparison

Although directional classification accuracy is the primary evaluation metric in this study, additional performance measures strengthened model validation and ensured robustness (

Table 9. Summary of validation metrics for the three fuzzy modeling strategies.

Strategy	Accuracy	Balanced Accuracy	Macro-F1	Cohen’s Kappa	MAE (m)	RMSE (m)
S1—Localized	0.671	0.615	0.606	0.421	0.319	0.846
S2—Unified	0.640	0.583	0.569	0.368	0.367	0.918
S3—Behavioral	0.698	0.649	0.647	0.468	0.298	0.826

). Specifically, balanced accuracy, macro-averaged F1-score, and Cohen’s Kappa coefficient quantified classification performance under class imbalance and accounted for agreement beyond chance. In addition, continuous error measures, namely Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), were computed based on predicted and observed groundwater depths. These supplementary metrics provide a complementary quantitative assessment of prediction errors and allow examination of consistency between directional and numerical model performance.

The inclusion of additional validation metrics confirms the robustness of the comparative results obtained using directional classification accuracy. The behavioral clustering strategy (S3) consistently outperforms the localized and unified approaches across all metrics, achieving higher balanced accuracy, macro-F1 score, and Cohen’s Kappa, indicating improved classification performance beyond chance agreement. Moreover, S3 exhibits the lowest MAE and RMSE, demonstrating its superiority in continuous groundwater depth predictions. These results indicate that the behavioral clustering framework provides a more balanced and physically consistent representation of groundwater dynamics under data-limited conditions.

A comprehensive comparison of predictive performance across all 35 monitoring wells highlights systematic differences among the three fuzzy modeling strategies (

Table 10. Comparative performance summary across modeling strategies (35 wells). Values computed using identical validation procedures.

Strategy Performance	Local	Unified	Behavioral
Mean Accuracy	0.71	0.67	0.74
Best Performance Count	10 wells	2 wells	23 wells
Performance Range	0.61–0.80	0.58–0.76	0.60–0.85
Standard deviation	0.053	0.052	0.071
Wells where Behavioral > Local	w1 (+0.02), w3 (+0.23), w4 (+0.01), w5 (+0.17), w9 (+0.16), w10 (+0.09), w14 (+0.07), w15 (+0.01), w17 (+0.09), w19 (+0.16), w21 (+0.14), w24 (+0.02), w26 (+0.13), w27 (+0.02), w28(+0.09), w30 (+0.01), w31 (+0.02), w32 (+0.10), w33 (+0.12), w34 (+0.15), w35 (+0.06)
Wells where Local > Behavioral	w2, w6, w7, w8, w11, w12, w13, w16, w18, w25, w29
Wells where Unified > Local	w5 (+0.11), w20 (+0.02), w22 (+0.03), w23 (+0.11)

; Figure 18). The behavioral clustering approach achieved the highest overall mean classification accuracy (0.74), followed by the localized (0.71) and unified (0.67) strategies. These differences indicate that incorporating behavioral similarity and additional temporal information can improve groundwater level prediction in heterogeneous systems.

Beyond mean accuracy, Table 10 and Figure 18 reveal important contrasts in performance variability. Behavioral clustering exhibits the broadest range of accuracy values (0.60–0.85) and the highest standard deviation (0.071), indicating that this strategy can achieve substantial accuracy gains at many wells but may also perform less effectively at locations with more complex or transitional behavior. This variability reflects the adaptive nature of behavior-based grouping, which emphasizes functional similarity rather than spatial proximity. The localized strategy shows intermediate performance characteristics, with a moderate mean accuracy and a narrower performance range (0.61–0.80) than behavioral clustering. Its lower standard deviation (0.053) suggests more stable performance within hydrogeographical groups; however, the wide distribution indicates that spatial grouping does not consistently capture dominant groundwater controls at all wells. As a result, localized models perform well at some locations but offer limited improvement at wells with atypical or mixed response behavior. The unified strategy displays the lowest mean accuracy but the smallest standard deviation (0.052), demonstrating the most consistent performance across all wells. This narrow distribution reflects the use of a single basin-wide rule base, which reduces extreme variability but also constrains the ability to represent local response differences. While unified models rarely achieve the highest accuracy at individual wells, they provide stable and predictable performance that may be advantageous for basin-scale assessments and preliminary screening applications.

The boxplots provide additional insight into the distributional characteristics of prediction accuracy across wells for each strategy. The median value represents the typical well-level performance, while the interquartile range (IQR) reflects the variability among wells. The behavioral clustering strategy exhibits higher median accuracy and a more compact IQR than the localized and unified models, indicating greater robustness and reduced sensitivity to site-specific heterogeneity. In contrast, the wider spread and longer whiskers observed for the localized strategy reflect stronger dependence on local hydrogeological conditions, leading to increased inter-well variability. The unified model shows the narrowest distribution but at a lower median accuracy, highlighting the trade-off between consistency and predictive precision.

Analysis at the individual well level further clarifies these trade-offs. Wells where behavioral clustering outperforms localized models represent locations where spatial or geological grouping does not adequately capture groundwater response dynamics. Wells where localized models outperform behavioral clustering typically reflect strong local geological control, where spatial grouping remains effective. The small number of wells where the unified strategy outperforms localized models indicates that basin-wide generalization offers limited but site-specific advantages at locations where groundwater responses closely follow regional climatic patterns.

The comparison reveals that no single modeling strategy is universally optimal. Predictive performance depends on the extent to which spatial structure, basin-wide climatic forcing, or functional response behavior govern groundwater dynamics. These results emphasize the importance of selecting fuzzy modeling strategies based on both the hydrogeological conditions of the study area and the intended application scale.

4.3.2. Performance Trade-Offs

Analysis of individual well performance highlights clear trade-offs among the three fuzzy modeling strategies (Figure 19). Differences in predictive accuracy reflect contrasting assumptions about local specificity, basin-wide generalization, and functional similarity in groundwater response. As a result, strategy effectiveness varies depending on the dominant controls governing groundwater dynamics at each location. Wells that achieve high accuracy under the localized strategy generally maintain comparable or slightly improved performance under behavioral clustering, whereas they show reduced accuracy under the unified approach. This pattern indicates that when local geological conditions strongly influence groundwater responses, spatially based grouping remains effective, and additional generalization may reduce predictive precision.

Wells that perform poorly under localized models often show substantial improvements in accuracy under behavioral clustering. This behavior suggests that functional grouping based on observed groundwater response characteristics can successfully capture relationships that hydrogeographical classification alone may not represent well. The inclusion of lagged groundwater level changes and seasonal indicators appears particularly beneficial for wells exhibiting memory effects and complex temporal dynamics.

The unified strategy demonstrates a different trade-off. Although it rarely achieves the highest accuracy at individual wells, it provides stable, consistent performance across the basin. This consistency may be advantageous for regional-scale applications where model simplicity, transferability, and uniform implementation are prioritized over local optimization. The performance trade-offs illustrated in Figure 19 confirm that no single modeling strategy is optimal under all conditions. Localized models favor site-specific accuracy, unified models favor consistency and scalability, and behavioral clustering provides a balance between adaptability and generalization. These trade-offs directly reflect the underlying assumptions of each strategy and should guide strategy selection according to the hydrogeological context and intended application scale.

To visualize these trade-offs, Figure 20 presents a direct time-series comparison for three representative wells (w19, w22, and w10). The panels illustrate how predictive performance evolves across the three strategies, highlighting differences not only in point-wise accuracy but also in prediction uncertainty. The Behavioral Strategy (bottom panel) shows the closest alignment with observed groundwater fluctuations while exhibiting narrower prediction intervals during key recharge and recession periods. This behavior indicates an improved ability to capture both the magnitude and timing of groundwater responses, effectively balancing site-specific precision with basin-wide generalization.

The time-series comparisons further clarify the differences in model behavior beyond point-wise accuracy metrics. The localized strategy captures site-specific fluctuations but may exhibit phase shifts or amplitude mismatches during recharge and recession periods. The unified strategy produces smoother predictions with reduced variability, reflecting basin-wide generalization but limited responsiveness to local dynamics. In contrast, the behavioral clustering strategy aligns more closely with observed groundwater fluctuations in both timing and magnitude, particularly during recharge events and seasonal transitions. The narrower prediction intervals indicate reduced uncertainty and improved representation of groundwater memory and delayed recharge processes.

In addition to predictive performance, the three modeling strategies differ in their computational and structural complexity. The localized strategy requires the construction and maintenance of multiple group-specific fuzzy rule bases, increasing setup effort and reducing scalability as the number of hydrogeographical groups grows. The unified strategy represents the simplest configuration, relying on a single basin-wide rule base with minimal computational and implementation overhead, but at the expense of reduced predictive accuracy. The behavioral clustering strategy introduces additional computational steps during the clustering and feature extraction phase; however, this cost is incurred only once during model setup. Importantly, the subsequent fuzzy inference remains lightweight, and the number of rule bases is reduced compared to the localized approach. The observed accuracy gains, therefore, outweigh the modest increase in initial computational effort, particularly in data-limited settings that prioritize improved generalization and transferability over minimal model complexity.

4.4. Time-Series Validation and Uncertainty Assessment

To explicitly address temporal autocorrelation and prevent potential data leakage, a time-series validation framework based on rolling-origin evaluation was implemented. This approach assessed model performance using an expanding training window followed by sequential testing windows. This method ensured predictions based on future observations beyond the training period. This validation scheme provided a more realistic assessment of model generalization under non-stationary groundwater conditions.

Figure 21 illustrates the distribution of directional classification accuracy across individual wells for the three modeling strategies. The behavioral clustering-based strategy (S3) exhibits higher median accuracy and a more compact interquartile range than the localized (S1) and unified (S2) strategies, indicating improved robustness and reduced sensitivity to site-specific variability.

To quantify prediction uncertainty, 95% confidence intervals (CIs) were estimated at the well level using a moving-block bootstrap procedure, thereby preserving the temporal dependence inherent in groundwater time series. As shown in Figure 22, the behavioral strategy achieves the highest mean accuracy, with confidence intervals that remain consistently above those of the alternative strategies, confirming the statistical robustness of its performance gains.

Rolling-origin cross-validation further examined the temporal stability of model performance. Figure 23 presents basin-averaged test accuracy across successive validation folds. While all strategies exhibit a gradual decline in accuracy with increasing forecast horizon—reflecting the effects of hydroclimatic non-stationarity—the behavioral clustering framework consistently outperforms the other approaches across all folds. This behavior suggests that grouping wells based on shared dynamic response characteristics enhances temporal transferability and mitigates performance degradation over time.

From a physical perspective, the observed temporal decline in accuracy is consistent with delayed recharge processes, seasonal variability, and memory effects characteristic of shallow coastal aquifers. The superior performance of the behavioral strategy indicates that clustering wells based on their dynamic groundwater responses implicitly captures these processes, thereby improving temporal generalization.

Beyond overall accuracy, the robustness of the behavioral strategy (S3) was further evidenced by its class-wise accuracy, achieving 71.45% for the ‘Rise’ category and 70.96% for the ‘Decline’ category. These complementary metrics indicate that the model’s performance is balanced across different hydrological states, ensuring reliable forecasting for both recharge and depletion phases.

5. Discussion

Although groundwater modelers have applied fuzzy logic, the advancement proposed in this study does not lie in modifying fuzzy inference mechanisms themselves, but in the implementation strategy and system organization. Specifically, the proposed framework advances existing applications by systematically evaluating alternative fuzzy grouping logics, integrating behavior-based clustering derived from observed groundwater responses, and adopting a directional classification paradigm aligned with decision-making needs under data-limited conditions. This implementation-level refinement enhances model interpretability, internal consistency, and transferability without increasing structural complexity, thereby extending the practical applicability of fuzzy expert systems beyond conventional single-structure implementations.

5.1. Methodological Insights

The comparative evaluation confirms that fuzzy expert systems provide an effective framework for groundwater forecasting in data-limited environments. Classification accuracies across the three modeling strategies range from 0.67 to 0.74, corresponding to 102–124% improvements over random classification (0.33). These performance levels are consistent with previous fuzzy-logic applications in hydrological modeling, where researchers report accuracies of 0.65–0.80 for complex environmental systems [30,33]. The results support the suitability of fuzzy approaches for representing climate–groundwater relationships through expert knowledge while accommodating uncertainty and nonlinearity. The behavioral clustering strategy demonstrates clear methodological advantages, achieving the highest mean accuracy (0.74) and outperforming the other strategies at 66% of the monitoring wells. The substantial improvement in accuracy observed at well w3 (from 0.62 to 0.85) illustrates the ability of behavior-based grouping to identify functional similarities that spatial or geological classification may not capture. The inclusion of lagged groundwater level changes and seasonal indices in cluster-specific rule bases improves the representation of temporal dynamics, particularly for wells exhibiting memory effects. This finding is consistent with previous studies highlighting the importance of temporal dependencies in groundwater systems [10,12].

Analysis of spatial grouping reveals limitations of hydrogeographical classification as the sole basis for model organization. Performance variability within hydrogeological groups ranges from 0.14 to 0.19, with Group C exhibiting the most extensive spread (0.61–0.80) and Group E the smallest (0.72–0.79). These variations indicate that geological classification alone does not fully explain groundwater response behavior. The weak relationship between spatial proximity and model performance (R² < 0.3) further suggests that functional behavior provides a more appropriate basis for specialized modeling than spatial or administrative boundaries. Studies of complex aquifer systems that emphasize hydrological function over spatial organization report similar conclusions [31,37].

Behavioral clustering improves internal consistency by grouping wells with similar temporal autocorrelation and climatic sensitivity, independent of location. The standard deviation of accuracy within behavioral clusters (0.06–0.09) is consistently lower than within hydrogeographical groups (0.08–0.14), indicating greater homogeneity in response behavior. This methodological outcome addresses a key limitation of traditional groundwater modeling approaches, where spatial grouping does not adequately capture dominant hydrological processes [21,26].

The increased structural complexity associated with behavioral clustering raises the potential for overfitting. In the present framework, several design choices mitigate this risk. First, the analysis limited the number of behavioral clusters to three, balancing model expressiveness with parsimony. Second, fuzzy rule bases were derived from representative wells and subsequently transferred to independent wells within each cluster, providing an implicit form of out-of-sample evaluation. Third, using directional classification rather than continuous regression reduces sensitivity to noise and local fluctuations. Finally, the stability of clustering results under different lag orders and seasonal representations further indicates that the behavioral grouping captures robust hydrological patterns rather than site-specific artifacts. Collectively, these elements limit overfitting while preserving the interpretability and generalization capacity of the behavioral clustering strategy.

Regarding transferability, it is important to note that the proposed behavioral clustering provides a methodological framework rather than a fixed set of clusters. While the specific cluster compositions and fuzzy rule bases are basin-dependent, the underlying feature-based clustering approach may work with other basins with different hydroclimatic conditions, provided that dominant groundwater response mechanisms and data availability are comparable.

5.2. Hydrogeological Consistency of Behavioral Clusters

Although groundwater-level dynamics drove the behavioral clustering strategy, an additional analysis assessed consistency between the resulting functional similarity and the hydrogeology. Since direct measurements of key hydrogeological parameters, such as hydraulic conductivity and porosity, were not available for all monitoring wells, lithological units, total well depth, and elevation were used as physically meaningful proxies.

Table 11. Hydrogeological characteristics of wells grouped by behavioral cluster.

Behavioral Cluster	Dominant Lithology (%)	Mean Total Depth ± SD (m)	Mean Elevation (Z) ± SD (m)	No. of Wells
Cluster 1	Q3 (40.0%)	25.2 ± 6.8	18.6 ± 3.5	10
Cluster 2	Q3 (53.3%)	30.1 ± 11.4	22.1 ± 7.2	15
Cluster 3	Q4 (60.0%)	25.5 ± 4.7	23.2 ± 8.6	10

summarizes a quantitative comparison of these parameters within and between behavioral clusters.

The results indicate that wells within the same behavioral cluster exhibit greater lithological homogeneity than those in different clusters. In particular, clusters dominated by Quaternary formations (Q4–Q3) contain shallower wells and exhibit lower elevation variability. In contrast, clusters containing older formations (Q2 and Neogene units) tend to include deeper wells and exhibit greater structural and vertical heterogeneity (Table 11).

Furthermore, the variability in total well depth and elevation within individual clusters is consistently lower than that observed across clusters. This pattern suggests that wells with similar groundwater-level dynamics are also characterized by comparable hydrogeological settings, thereby providing quantitative support for the assumption that behavioral similarity reflects underlying physical controls rather than purely statistical coincidence.

From a physical standpoint, this correspondence is consistent with the expected behavior of coastal aquifer systems, in which younger Quaternary deposits typically exhibit higher effective permeability and a more rapid hydraulic response to recharge. At the same time, older, more consolidated units respond with delayed and attenuated groundwater dynamics. Overall, the hydrogeological consistency highlighted in Table 11 provides independent physical support for the behavioral clustering framework and further strengthens the interpretability and robustness of the proposed modeling strategy.

5.3. Practical Implications

The results provide clear guidance for selecting modeling strategies under different management objectives and resource constraints. Localized models achieve optimal performance at 29% of wells. They are most appropriate for targeted management of specific sub-areas where high accuracy is required and sufficient site-specific data are available. The strong performance at wells w7 (0.79), w12 (0.80), and w16 (0.79) illustrates the effectiveness of localized models when hydrogeological conditions align well with spatial grouping assumptions. Behavioral clustering supports regional groundwater management applications by achieving the best performance at 66% of wells while requiring only three distinct rule bases, compared to five in the localized strategy. This reduction in model complexity, combined with higher mean accuracy, makes behavioral clustering suitable for managing heterogeneous groundwater systems with moderate computational resources. The improvement observed at 21 wells, each showing an accuracy gain of 0.05 or more, demonstrates the method’s ability to capture groundwater response behavior that spatial classification cannot. Unified models maintain practical relevance despite lower mean accuracy (0.67). Their consistent performance across hydrogeological groups, reflected by the lowest standard deviation (0.052), supports their use in preliminary assessments, ungauged locations, and resource-limited applications. Although the unified strategy achieves the highest accuracy at only 6% of wells (w20, w22, w23), it produces acceptable accuracy (>0.60) at 89% of sites, supporting its role as a baseline modeling approach [27,38]. Behavioral clustering requires additional computational effort for feature extraction and clustering, estimated at approximately 15% more processing time than unified models. This additional effort results in an 11% increase in mean accuracy and a 23% reduction in poorly performing wells (accuracy < 0.65). Organizations with extensive monitoring networks and analytical capacity can therefore benefit from improved performance, while simpler strategies remain viable where technical resources are limited.

Implementation considerations include data availability, with a minimum of 24 months of synchronized records required for reliable behavioral classification, as well as technical capacity for statistical and GIS analysis. The modular structure of fuzzy expert systems supports incremental adoption, allowing implementation to begin with unified models and expand to behavioral clustering as data availability and expertise increase [29].

Although direct comparison with artificial intelligence and deep-learning models was beyond the scope of this study, the reported classification accuracies (0.67–0.74) are comparable to, and in some cases within the lower-to-mid range of, values reported for data-driven groundwater forecasting models in semi-arid environments. Many AI-based approaches achieve higher numerical accuracy but rely on large training datasets and operate as black boxes, limiting interpretability and transferability. In contrast, the proposed fuzzy framework emphasizes decision-oriented performance, transparency, and robustness under data-limited conditions. Accordingly, the performance gains reported here should be interpreted not as absolute numerical superiority, but as a practical improvement within a modeling paradigm specifically designed for data-scarce groundwater systems.

6. Conclusions

This study developed and evaluated three fuzzy modeling strategies for predicting monthly changes in groundwater levels in the semi-arid Al-Hsain Basin, Syria, using 4 years of data (2020–2024) from 35 monitoring wells. The strategies included localized fuzzy models, a unified basin-wide model, and a behavioral clustering-based approach. The comparative analysis revealed differences in predictive performance among the three methods. The behavioral clustering strategy achieved the highest overall mean classification accuracy (0.74), outperforming localized models (0.71) and the unified model (0.67). It achieved the best performance across 66% of wells and produced substantial improvements at locations where other strategies performed poorly, with accuracy gains of up to 0.23 per well. By grouping wells based on response behavior rather than spatial or geological similarity, this approach identified functional hydrological patterns that transcend traditional hydrogeographical boundaries. Behavioral clustering also reduced intra-group performance variability (0.06–0.09) compared to geographical grouping (0.08–0.14), indicating improved internal consistency.

Localized models performed best at 29% of wells, particularly where aquifer conditions aligned well with spatial and geological assumptions. Group-level mean accuracies ranged from 0.68 to 0.76. However, notable variability within hydrogeographical groups highlighted the limitations of relying solely on spatial organization in complex groundwater systems. The unified basin-wide model showed the most consistent performance across the study area, achieving acceptable accuracy (>0.60) at 89% of wells. While it rarely achieved the highest accuracy at individual locations, its stability and low data requirements make it suitable for preliminary assessments and applications with limited resources. The inclusion of lagged groundwater level changes and seasonal indices in the behavioral clustering models was critical for capturing temporal complexity and groundwater memory effects. This input structure reduced the number of required specialized rule bases from five to three while maintaining superior predictive performance. Overall, the results show that optimal strategy selection depends on application objectives, data availability, and required accuracy. Behavioral clustering offers a balanced solution for regional groundwater management, improving mean accuracy by 11% and reducing model complexity by 40% compared to traditional spatial organization. The approach requires at least 24 months of synchronized data and approximately 15% additional computational effort. However, it yields substantial gains in predictive reliability, especially in heterogeneous aquifer systems where spatial proximity correlates weakly with hydrological behavior.

Further improvements may incorporate additional input variables, such as evapotranspiration, soil moisture, and groundwater abstraction data, to better represent the water balance in semi-arid environments. Extending the framework to higher temporal resolution (weekly or daily forecasting) could enhance early warning capabilities for groundwater decline. Hybrid approaches combining fuzzy logic with machine learning, such as ANFIS and ensemble models, may further improve robustness while preserving interpretability. Evaluating the transferability of behavioral clustering across different hydroclimatic regions would help define broader applicability and implementation guidelines.

This study shows that behavior-based clustering of groundwater monitoring wells provides a reliable and interpretable framework for groundwater-level forecasting in data-limited, hydrogeologically complex settings. The approach offers a practical basis for supporting groundwater management decisions under increasing climatic variability and water demand.