Towards interpreting machine learning models for predicting soil moisture droughts

Abstract Determination of the dominant factors which affect soil moisture (SM) predictions for drought analysis is an essential step to assess reliability prediction results. However, artificial intelligence (AI) based modelling only provides results without physical interpretation models. Here, we propose explainable AI (XAI) framework reveal SM events. Random forest site-specific models were developed using data from 30 sites, covering 8 vegetation types. The unity multiply XAI tools was applied interpret site-models both globally (generally) and locally. Globally, interpreted two methods: permutation importance accumulated local effect (ALE). On other hand, each event, locally via Shapley additive explanations (SHAP), interpretable model-agnostic explanation (LIME) individual conditional expectation (ICE) methods. features identified as temperature, atmospheric aridity, time variables latent heat flux. But through interpretations events, showed a greater reliance on aridity flux at grass with higher correlation time-dependent parameters sites located in forests. temporal variation feature effects events also demonstrated. could shed light how are made promote application techniques prediction, may be useful irrigation water resource management.