On the improved ensemble of multi-source precipitation through joint automated machine learning-based classification and regression

Accurate precipitation representation at local and global scales will greatly improve our understanding of climate system changes. However, no precipitation estimate consistently has the lowest errors (systematic biases, random error, and rain/no-rain classification error) under varying environmental gradients, resulting in considerable uncertainty when investigating mechanisms and making predictions. Multiple Source Precipitation Ensemble (MSEP) is regarded as an indispensable approach to this challenge. Based on an automatic machine learning workflow, we propose an MSPE framework that uses machine learning classification and regression jointly. Six distinct precipitation products (e.g., satellite- and reanalysis-based estimates) and their ensembles based on different framework strategies were examined comprehensively at 818 gauges across China and 500 randomly selected sites (representing ungauged regions). The unique features of MSPE were investigated, including the necessity of assigning spatiotemporal dynamic weights and the usage of machine learning classification and regression jointly. Results demonstrated that MSPE could effectively reduce both random and classification errors associated with precipitation occurrences. In addition, the capacity to generalize and the interpretability of the ML models developed within the framework were compared and discussed in depth. We also summarized the current framework's limitations and potential expansions. The framework presented in this research is expected to be a robust and flexible framework for the global application of ensembles of precipitation estimates from numerous scales, data sources, and time periods.