| Summary: | Precipitation estimates with high accuracy and fine spatial resolution play an important role in the field of meteorology, hydrology, and ecology. In this study, support vector machine (SVM) and back-propagation neural network (BPNN) machine learning algorithms were used to downscale the Integrated Multi-satellite Retrievals for Global Precipitation Measurement (GPM) (IMERG) data at daily scale through four events selected from 2017 and 2018 by establishing the relationships between precipitation and six environmental variables over Zhejiang, Southeastern China. The downscaled results were validated by ground observations, and we found that (1) generally, the SVM-based products had better performance and finer spatial textures than the BPNN-based products, the multiple linear regression (MLR)-based products, and the original IMERG; (2) all downscaled products decreased the degree of overestimation of the original IMERG at heavy-precipitation regions to a certain extent; (3) for heavy-precipitation events in the plum rain season, the downscaled products based on SVM and BPNN both improved prediction accuracy compared to the MLR-based products and the original IMERG considering the validations against ground observations. R2 maximally increased from 0.344 to 0.615 for the SVM-based products and from 0.344 to 0.435 for the BPNN-based products compared to the original IMERG; and (4) for typhoon precipitation events, the SVM-based products still showed better accuracy with R2 maximally increased from 0.492 to 0.615 compared to the original IMERG. In contrast, the performance of BPNN-based products was not satisfying and showed no significant differences with the performance of MLR-based products. This study provided a potential solution for generating downscaled satellite-based precipitation products at meteorological scales with finer accuracy and spatial resolutions.
|
|---|
