Causes of Watershed Drought Analyzed Using Explainable Deep Learning: A Case Study in Fenhe River Basin

Zixuan Chen¹Xikun Wei^2*Guojie Wang³Yifan Hu³Haonan Liu³Jinman Zhang¹Shuang Zhou¹Zengbao Zhao¹Yushan Liu¹

• ¹Key Laboratory of Meteorology and Ecological Environment of Hebei Province, Meteorological Institute of Hebei Province, Shijiazhuang, China
• ²Collaborative Innovation Center for Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science & Technology, Nanjing, Jiangsu Province, China
• ³School of Remote Sensing & Geomatics Engineering, Nanjing University of Information Science and Technology, Nanjing, Jiangsu Province, China

The study predicted daily-scale drought for the Fenhe River (FHR) basin and applied the explainable artificial intelligence (XAI) method to the model's prediction results. Daily-scale drought prediction can provide more timely and detailed drought information, while deep learning interpretable methods can help to understand the impact of different predictors on drought and improve the credibility of the model. The Standardized Antecedent Precipitation Evapotranspiration Index (SAPEI) was selected as an index for evaluating drought conditions. Five classical deep learning prediction models including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional Long Short-Term Memory Networks (biLSTM), Transformer (TFR), Informer (IFR) were applied in the experiment, and the performance of each model was comprehensively evaluated. The results of the test set show that all models make effective predictions of drought in the FHR basin with pearson correlation coefficient (R) higher than 0.75. BiLSTM performs better in short-term prediction, while TFR and IFR are better at long-term prediction. The results of the deep learning interpretable model show that putting aside the strong influence of SAPEI itself in the prediction process, mean temperature (TM) has the greatest influence among the auxiliary predictors, followed by precipitation (PRE) and relative humidity (RHU), with potential evapotranspiration (PET) being the weakest. Our work emphasizes the importance of timely warnings of drought and the importance of XAI for the development of artificial intelligence.