What’s Behind This Water Table Depth Forecasting? RISE Application for Spatial, Temporal, and Spatio-Temporal Explanations

Over the past few years, the effects of climate change have significantly increased, directing our attention to our environmental resources. One of the most critical resources is fresh water, whose availability has been endangered by even more frequent extreme events of precipitations and droughts. Deep Learning (DL) has revealed a useful tool for accurate hydrological forecasting, but its main drawback is its black-box nature. To face this issue eXplainable Artificial Intelligence (XAI) has come out. In this work, Randomized Input Sampling for Explanation (RISE) is applied to a regression spatio-temporal model trained to predict the Water Table Depth (WTD) collected by a sensor in Vottignasco, in the northwest of Italy. An investigation of the model behaviour over spatial, temporal and spatio-temporal dimensions has been conducted formalizing S-RISE, T-RISE, and ST-RISE respectively. Results suggest the usefulness of a spatio-temporal approach (ST-RISE) and give interesting intuitions on the events that occurred over the area.