Soil infiltration is a critical process shaping water security in Ethiopia’s northwestern highlands, determining whether rainfall nourishes crops, recharges aquifers, or is lost as runoff. In the Debre Markos region, infiltration varies widely with land use, soil texture, and seasonal rainfall, yet many models still assume uniform behaviour. To address this gap, we developed a two-stage hybrid framework that integrates the physically based Green-Ampt model with Support Vector Machine (SVM) learning. The Green-Ampt model provided baseline infiltration estimates using calibrated soil hydraulic parameters (RMSE = 23.1mm/hr, R² = 0.73), while SVM regression corrected residual biases. The (RMSE = 19.5mm/hr, R² = 0.79, representing a 15.6% error reduction) and SVM classification delineated infiltration zones into four categories based on steady-state infiltration rates. These thresholds were selected to align with Ethiopian highland soil and land-use management priorities, where very low rates indicate severe compaction or surface sealing, and high rates correspond to well-vegetated, permeable soils. Results revealed that grasslands on silty loam exhibit high point-scale infiltration capacity (>300mm/hr), while urban and grazed lands on clay loam show suppressed rates (<150mm/hr), indicating their relative potential for local water entry rather than direct groundwater recharge. These findings demonstrate how combining physical models with AI enhances predictive reliability and provides a reproducible framework for identifying recharge-prone areas and erosion-vulnerable zones, supporting site-specific watershed management and climate adaptation in Ethiopia’s highlands.
Reliable daily streamflow prediction is essential for water resources management in California’s Sierra Nevada, where increasing hydroclimatic variability has intensified flood and drought risks. Although regional machine learning models can improve generalization across watersheds, conventional shared architectures may not adequately capture watershed-specific characteristics and distinct flow regimes. To address this limitation, we propose MoE-τ, a mixture-of-experts framework that combines multiple LSTM experts through an attention-based temporal gating network with a learnable temperature mechanism for adaptive expert weighting. We evaluated MoE-τ across 14 heterogeneous Sierra Nevada watersheds against learning-based models and the SAC-SMA process-based model. Under the in-domain temporal prediction benchmark, MoE-τ attained the highest mean NSE (0.873) among the learning-based models, a modest improvement over the strongest recurrent baseline (Bi-LSTM, 0.857) and a larger margin over SAC-SMA (0.762). It also remained robust under an alternative train–test split. Ablation and parameter-scale analyses showed that the performance gains arise from the proposed gating and expert-based design rather than increased model size. Model interpretation analyses revealed expert specialization across flow regimes and a strong relationship between hydrometeorological distribution differences and expert-usage differences (Spearman’s r=0.687). In held-out watershed experiments, MoE-τ outperformed LSTM on nearly all metrics across watersheds with distinct hydrologic conditions. These results demonstrate that adaptive expert selection improves both in-domain prediction and generalization to unseen watersheds, providing a robust and interpretable method for regional streamflow modeling.
Groundwater recharge remains a critical yet poorly quantified component of the hydrological cycle, particularly in data-scarce, heterogeneous basins. This study employs a novel comparative framework that integrates stable isotope mass balance with the spatially distributed WetSpass-M hydrological model to assess groundwater recharge dynamics across three climatically and topographically distinct watersheds (Gilgel Gibe, Gombora, and Neri) within the Omo Gibe River Basin, Ethiopia. We analyzed stable isotopes of Oxygen (δ18O) and Hydrogen (δ2H) in precipitation and groundwater to establish Local Meteoric Water Lines (LMWLs) and quantify seasonal recharge contributions. Concurrently, the WetSpass-M model was calibrated and validated against stream flow data to simulate high-resolution spatiotemporal components of the water balance. Key findings include: (1) groundwater recharge is primarily driven by wet-season precipitation, with isotopes revealing that 88.1% (Gilgel Gibe) and 83.7% (Gombora) of annual recharge occurs during the wet season; (2) in the bimodal Neri watershed, a more balanced seasonal distribution is observed (64.7% wet season), reflecting its distinct rainfall regime; and (3) the WetSpass-M model corroborates these seasonal proportions while providing high-resolution spatial mapping of recharge hotspots. The WetSpass-M model provided detailed spatial mapping, confirming that recharge is highest in vegetated highlands with permeable soils and is significantly reduced in urbanized areas. The strong agreement between the two methods underscores the robustness of our synergistic framework, in which isotope data validate the model's temporal dynamics and the model elucidates the spatial context of the isotopic signatures. This study suggests that sustainable groundwater management in the basin should prioritize wet-season water harvesting, land-use planning to protect infiltration zones, and climate adaptation strategies.
Floods and landslides are significant natural hazards that frequently threaten the Indian Himalayas, an extremely sensitive region both geologically and climatically. Shallow slope failure due to heavy rainfall during rainstorms is common in mountain areas. Existing forecasting models are usually either purely data-driven and thus suffer from the "black-box" issue and poor generalization, or purely physics-based, which are computationally expensive and extremely challenging to calibrate. This research presents a hybrid framework combining a Physics-Informed Neural Network (PINN) for flood forecasting and an Extreme Gradient Boosting (XGBoost) model for landslide susceptibility prediction. It serves as a compromise between purely data-driven and physics-based models. Our method is an integration of observed data consisting of high priority variables - precipitation records, soil moisture data, and Digital Elevation Models (DEMs) with the underlying governing partial differential equations (PDEs). By incorporating physical equations and laws into the neural network, the Hydrological model obtains proficiency such that the results align well with observed data and physical behavior, governed by the physical laws. The proposed method is applied in a highly dynamic Himalayan region (Himachal Pradesh), a region that frequently witnesses the two natural hazards.
Artificial groundwater recharge (AGR) is an effective method to replenish depleted aquifers in arid and semi-arid regions facing over-extraction, droughts, and rising demand. By channeling surface water underground (e.g., via infiltration basins or wells), AGR boosts storage, improves quality through soil filtration, and supports sustainable management in water-stressed areas. This research develops a novel framework integrating Ordered Weighted Averaging (OWA) and Principal Component Analysis (PCA) for optimal AGR site selection in water-scarce regions, using the Yazd-Ardakan watershed as a case study. From 17 geospatial layers, eight critical factors were identified (correlation >0.5): land use, slope gradient, proximity to aquifers, geological formations, pedological characteristics, elevation bands, temperature variations, and microclimate zones. PCA showed the first principal component (PC1) explaining 35.9% of variance, with dominant influence from pedological properties (weight=0.32), geological features (0.29), slope gradients (0.27), and elevation ranges (0.25). OWA and PCA outputs had a strong 64% positive correlation, validating these factors as key determinants for recharge suitability. The integrated approach reduces data processing by 30-40% compared to conventional methods, provides statistical validation of weights via PCA, and adapts well to arid/semi-arid environments. It offers water managers a cost-effective tool for sustainable AGR implementation and water security in climate-vulnerable regions.
