Irrigation water quality assessment and evaluation of machine learning prediction reliability in a semi-arid region, Midmar Catchment, Tigray, Northern Ethiopia

Groundwater is a key source of irrigation water in semi-arid regions, yet the reliability of machine learning (ML) predictions of irrigation water quality indices remains poorly understood, especially for small datasets often evaluated with a single train-test split. This study assesses groundwater suitability for irrigation in the Midmar catchment, northern Ethiopia, and evaluates the reliability of ML-based IWQI prediction using progressively rigorous validation schemes. Forty groundwater samples were analyzed for electrical conductivity (EC), total dissolved solids (TDS), total hardness (TH), sodium adsorption ratio (SAR), sodium percentage (Na%), residual sodium carbonate (RSC), permeability index (PI), Kelly index (KI), potential salinity (PS), and IWQI. Fifteen regression algorithms were evaluated across six input scenarios using single-split, repeated five-fold cross-validation, leave-one-out cross-validation (LOOCV), and nested cross-validation. EC (5%), TH (5%), and PI (37.5%) exceeded recommended limits, while TDS, SAR, RSC, Na%, KI, and PS remained within recommended limits. The IWQI classified 95%, 2.5%, and 2.5% of samples as no, low, and moderate restriction, respectively. Single-split and LOOCV evaluations indicated moderate predictive performance (SVR R² = 0.666; GPR R² = 0.694). However, repeated cross-validation reduced the maximum mean R² to 0.179, and nested cross-validation produced negative mean R² for all models and scenarios. Groundwater is generally suitable for irrigation, though elevated EC, TH, and PI at some locations warrant monitoring. The pronounced decline in ML performance under rigorous validation indicates that small datasets may yield unstable and potentially misleading performance estimates, highlighting the need for larger, multi-seasonal datasets and robust validation before ML-based IWQI prediction can be considered reliable.