Explainable Machine Learning for Simultaneous Prediction of Soil Texture and Gravimetric Water Content: A SHAP-Driven Multi-Model Framework

Abstract

 Soil texture and gravimetric water content (GWC) are important properties which need to be determined accurately for proper irrigation and sustainable land management. Although machine learning (ML) has enhanced the ability to predict pedometrics, many of the high performing algorithms have been described as a "black-box" and practitioners have been reluctant to trust the model and connect the model output with soil physics. Furthermore, the existing approaches mostly assume independence between the attributes to be predicted and do not take into account their physical relationships. The present study aims to fill these voids by proposing a robust multi-model approach for simultaneous prediction of soil texture and GWC. A variety of data was used to assess state-of-the-art algorithms, such as XGBoost, CatBoost, and Random Forest. However, to achieve local and global interpretability, we integrated Shapley Additive Explanations (SHAP). This XAI-based approach filled the gap between computational intelligence and soil physics, allowing a fine-grained diagnosis of the contribution of the different features. The results show that the Random Forest framework outperformed in terms of concurrent performance with R2 value of 0.908, accuracy of 90.8% and Cohen's Kappa of 0.82 for soil texture and R2 value of 0.798 for GWC. Two features, Cation Exchange Capacity (CEC) and Organic Matter Percentage (OMP), were identified as the most influential using SHAP analysis, and they were found to exert the greatest influence on the moisture retention thresholds, with CEC being the most discriminative feature between the Clayey and Sandy texture classes, which aligns with the model's validity in light of the known pedology. This transparency and simultaneous monitoring of key soil properties makes it an appealing solution for precision agriculture and for resource management in the context of climate change.