Impact of overlooked input structure on rainfall forecasting performances of LSTM and XGBoost models

Reliable monthly rainfall forecasting is a non-trivial task due to the inherent complexity and chaotic behavior of rainfall process. Recently, the efficiencies of long short-term memory (LSTM) and extreme gradient boosting (XGBoost) have been demonstrated in numerous rainfall forecasting studies. Nevertheless, no study has investigated the performance of these models for high-dimensional feature space, which can significantly improve forecasting performance. To this end, this study analyses the efficiencies of 12 filter and wrapper methods across three stations in the Antalya Basin, Türkiye, for the first time. The results indicated that the minimum redundancy maximum relevance, statistically equivalent signatures, and max-min parents and children methods exhibited relatively consistent and acceptable performance. Furthermore, LSTM yielded better performance in rainfall forecasting, and XGBoost showed a higher dependency to input structure while providing better uncertainty estimations. The findings could enhance the univariate rainfall forecasting for effective water resource management and strategies for vulnerability adaptation.