Automated machine learning for soil salinity prediction and explainable driver analysis in drip-irrigated cotton fields of arid regions

Accurate monitoring of soil salinity is essential for optimizing water-salt regulation strategies and ensuring sustainable agricultural production in drip-irrigated cotton fields in arid regions. Although machine learning has become an effective tool for characterizing complex water-salt transport processes due to its strong nonlinear mapping capabilities, traditional modelling approaches face the dual challenges of low efficiency in manual trial-and-error and difficulties in interpreting "black-box" mechanisms, making it difficult to simultaneously achieve global architectural optimization and effective physical mechanism analysis. To overcome these limitations, this study developed and systematically evaluated the performance of four automated machine learning (AutoML) frameworks, including AutoGluon, H2O AutoML, FLAML, and TPOT, in predicting soil salinity based on in-situ field measurement data. SHAP (Shapley additive explanations) and PDP (Partial dependence plot) analyses were used to reveal the driving mechanisms of soil salt transport and their nonlinear response characteristics to environmental factors. The results indicated that AutoGluon achieved the optimal balance between computational efficiency and generalization accuracy through its multi-layer stacking ensemble strategy. In the 0–20 cm surface soil layer where water-salt transport was intense, the R 2 on the validation set reached 0.91–0.96, with the RMSE as low as 0.80–1.18 g·kg −1 , effectively correcting the systematic underestimation bias observed in the other AutoML frameworks within the dataset-defined high-salinity range exceeding 10 g·kg −1 . Variable importance and SHAP analysis further demonstrated significant vertical differences in salt-driving mechanisms. Compared with H2O AutoML, which relied excessively on biological indicators such as stem diameter, AutoGluon accurately captured physical driving processes, showing that salt transport in the 0–20 cm soil layer was primarily influenced by irrigation volume and growth stage, with a combined contribution exceeding 50%, while the 60–80 cm deep layer transitioned to being meteorologically dominated, with the contribution of cumulative ET o increasing to 29%. PDP analysis quantified the nonlinear physical response boundaries of key environmental factors, objectively identifying cumulative ET o in the 420–500 mm range as the critical threshold for inducing deep layer water upward movement and root zone salt accumulation, while also revealing a significant diminishing marginal benefit of leaching as irrigation approached 600 mm. This study confirms that AutoGluon achieves the unification of accurate soil salinity prediction and physical interpretation, and the quantified dynamic environmental response intervals provide a reliable data-driven basis for formulating precision irrigation schedules to suppress salinization and optimize water use in drip-irrigated cotton fields in arid regions.