Prediction of soil moisture via feature selection, model optimization, and climate data integration

Abstract

Soil moisture is vital for water resource management and agricultural productivity. With the growing availability of reanalysis data and advanced machine learning methods, it is essential to explore reliable, scalable models for long-term estimation. Comparing different algorithms enables a comprehensive assessment of their strengths and limitations in modeling complex soil moisture dynamics under varying climatic and agricultural conditions. In this study, soil moisture data for the period 1950-2022 from six stations in the Konya Closed Basin of T & uuml;rkiye were estimated using Echo State Networks (ESN), Adaptive Boosting (AdaBoost), Deep Neural Networks (DNN), and Light Gradient Boosting Machine (LGBM). Precipitation (PR), solar radiation (RSDS), wind speed (WS), mean temperature (Tave), maximum temperature (Tmax), minimum temperature (Tmin), and potential evapotranspiration (ET0) values calculated by the Thornthwaite equation were used as input variables of the models, and soil moisture (SM) was determined as the output variable. Model performance was evaluated using statistical metrics and visualizations derived from training, validation, and test datasets. In addition, a feature selection approach was used to identify the most effective parameters for soil moisture estimation. The analysis showed that the DNN model achieved the highest accuracy, with R2 values ranging from 0.87 to 0.97. The second most successful model was AdaBoost or ESN, while the LGBM model showed higher deviations than the others. These findings reveal that artificial intelligence-based models can effectively estimate soil moisture.