A physics-informed machine learning approach to predict soil water content for agricultural decision-making

This paper presents a novel hybrid modeling approach that combines physics-based modeling and ma-chine learning (ML) methods to enhance soil water con-tent dynamics prediction for agricultural decision-making. The proposed approach outperforms current methods for predicting water content based on soil-water physics or purely on data-driven strategies. Initially, a Markov chain-based model is employed to estimate soil water content. The uncertainty associated with the model-based estimation is then quantified by applying ML algorithms, such as support vector machine (SVM), random forest (RF), and feedforward neural network (FNN), to the soil water content data obtained from the Kansas mesoscale network (Mesonet). This quantified uncertainty reveals the complex soil water content dynamics that analytical Markov chain-based models cannot capture. The term physics-informed ML (PIML) refers to integrating soil-water physics principles with the predictive capabilities of data-driven models. Furthermore, multiple Mesonet time series datasets are utilized to assess the influence of physics-based knowledge on ML predictions. The evaluation of the proposed PIML models demonstrates significant improvements in predicting soil water content.