Performance Assessment of Bayesian Networks for Soil Moisture Prediction in Agricultural Water Management

Abstract

Estimating soil moisture using easily measurable parameters can significantly improve irrigation management efficiency while reducing time and cost. This study evaluated the performance of LASSO, Decision Tree, and Random Forest models for predicting soil moisture. In addition, Bayesian networks were used to model the complex interactions between soil, weather, and topographical factors at the Schäfertal site in Germany. To achieve this, seven structure learning algorithms, including PC, Grow-Shrink, Incremental Association, Fast Incremental Association, Hill Climbing, Max-Min Hill-Climbing, and Tabu Search, were used. Subsequently, soil moisture was predicted using all these algorithms to determine the best performance among them. The results showed that Bayesian networks had the best performance for prediction, while LASSO, Decision Tree and Random Forest did not perform as well, likely due to the complex relationships between variables. Bayesian networks, however, effectively identified key factors like soil texture, depth, and evapotranspiration using measured variables across all search algorithms. Bayesian networks showed better performance than the other three models. More importantly, they predicted soil moisture using only a subset of variables, with performance close to that achieved using the full set of variables. This result demonstrates improvements in predictions and handling missing data in practical applications such as precision agriculture and water management