Explainable AI – based study of the interactions between remote sensing and ground-truth climate variables and Lake Chad’s level fluctuations

Abstract - Study area: Lake Chad (Republic of Chad). - Purpose: Finding the significant remote sensing and ground-truth climate factors; their contributions and interactions in predicting remote sensing and ground-truth lake levels. Linear model (LM), regression tree (RT), random forest (RF), and gradient boosting regression (GBR) are employed. The best model is interpreted using Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). As results, GBR outperforms in regressing both remote sensing and ground-truth data. Regressing ground-truth lake level on ground-truth features outperforms regressing ground-truth lake level using remote sensing features. GBR-based LIME and SHAP analyses reveal that ground-truth air temperature influences the most ground-truth lake level; higher temperatures decrease the prediction, lowers increase it. Higher ground-truth precipitations decrease the prediction; lower precipitations increase it. Higher ground-truth evaporations and values within a specified range favor the prediction; lowers disfavor it. Remote sensing precipitation impacts the most ground-truth lake level; higher precipitations decrease the prediction; lowers and amounts within a specified range increase it. Higher remote sensingevaporations increase the prediction; lowers and values within a specified range reduce it. Higher remote sensingair temperatures decrease the output; lowers and values within a specified range favor it. Air temperature is the topmost factor either it is remote sensing or ground-truth. Precipitation and evaporation are 90% clustered either they are remote sensing or ground-truth. These results are additional essential tools for climate change impact and water resources management decision-makers in the region. More studies are needed for validation and application purposes.