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

Several severe drought events occurred in the past years and droughts will likely occur more frequent and be more intense in the future. Hydrological drought, which reflects the shortage of water in the river system, can lead to economic losses and can have severe negative impacts on aquatic ecosystems. Therefore being able to predict and increase insights in which rivers are more vulnerable to hydrological droughts, based on catchment characteristics and human interactions, can be of relevance for water managers. In this analysis, the drought sensitivity of rivers is predicted at a regional scale (Flanders, Belgium). Hereby the interests of multiple stakeholders is taken into account by considering four drought metrics, namely the yearly summer volume, the number of dry days, the drought intensity and drought severity. Whereby the latter three are based on the ecological flow. To predict each of these drought metrics, five models ranging from statistical to tree-based methods are applied using twelve input variables ranging from catchment characteristics to human interactions. Hereby random forest without bootstrap and XGBoost outperforms the other methods. To increase the interpretability of the results, the XGBoost models are used to calculate the SHAP and SHAP interaction values. As a result, the impact of the different input variables on the model results is assessed.

From this analysis, some general conclusions can be drawn. Irrigation is the most important variable for each of the considered drought metrics. However, not for every drought metric a clear, unique dependence between the irrigation and the drought sensitivity of a river could be observed. Rivers which have sand as dominant soil texture in their drainage area are less vulnerable to drought. When there are more human interaction in the drainage area, the river is more vulnerable to drought. Beside this, several other dependencies are observed of which many can be explained by the difference in ease of water transferability between sandy soils and clay soils. Next to this, it became clear that the impact of forest and agricultural area on the drought sensitivity of a river is complex, whereby especially its interaction with soil texture and human activities needs further investigation. The applied method can predict the drought sensitivity of a river based on catchment characteristics and human interactions, and therefore define rivers that are more vulnerable to drought. They moreover can provide additional insights in the importance of catchment characteristics and human interactions, and their relation to the drought sensitivity of a river.