An Interpretable Machine Learning Approach for Quantitative Precipitation Estimation from Multi-Source Remote Sensing Data

Machine learning (ML) models have become powerful tools for meteorological applications, but many still lack the transparency needed for scientific interpretation. This study addresses this issue by applying the interpretable ML approach RuleFit to quantitative precipitation estimation (QPE). The model was developed using a multisource dataset from eastern China, including composite radar reflectivity from China’s New Generation Doppler Weather Radar, Digital Elevation Model data, and Himawari-8 satellite Bands 7–10. A geographical mask based on China’s national borders was used to exclude points outside the land area, and standard scaling was applied to all input features to account for differences in units and value ranges. RuleFit was evaluated against six baseline models and the traditional Z-R relationship. The results show that RuleFit provides competitive predictive performance rather than a uniform advantage across all metrics. On the resampled dataset, it achieved a CSI (critical success index) of 0.602 and a POD (probability of detection) of 0.936 for 10-minute precipitation events exceeding 2 mm, which is comparable to other ML models and clearly better than the traditional Z-R relationship. Additional binary verification at multiple precipitation thresholds further shows that RuleFit maintains a reasonable balance between detection capability and false-alarm control across different rainfall intensity categories, although it is not consistently the top-performing model at every threshold. More importantly, the sparse linear terms, explicit decision rules, and partial dependence plots produced by RuleFit provide transparent insights into the model logic, revealing physically meaningful nonlinear relationships among radar, satellite, and terrain features. These findings suggest that the main value of RuleFit for QPE lies in combining competitive predictive skill with direct interpretability, making it a useful framework for both precipitation estimation and scientific understanding.