Analysing spatial variability in drought sensitivity of rivers using explainable artificial intelligence

Abstract

Hydrological drought can have a severe negative impact on aquatic ecosystems and human living standards. Therefore, being able to predict and gain more insights in the spatial variability in drought sensitivity of rivers is of relevance for water managers. The drought sensitivity of a river is in this study represented by four drought metrics, of which three are relative towards the ecological minimal flow. Statistical and machine learning methods were evaluated to predict these metrics for rivers in the Flanders region of Belgium based on catchment characteristics and data on human interferences. XGBoost had the best performance, with an explained variance of 80 % to 90 %. After applying explaining AI on these models, insights were obtained in the spatial variability of the drought metrics. Irrigation is the most important variable, a high percentage of irrigation leads to a higher drought sensitivity. If there are a lot of human interferences, there is a higher drought sensitivity. Many of the observed dependencies can be explained by the differences in soil infiltration capacity and transferability of water for sandy versus clay soils. No clear dependence with the amount of forest or agriculture was observed, implying that the impact of forest and agriculture on the drought sensitivity of a river is complex.