Application of Machine Learning Techniques in Satellite Precipitation Detection Using Himawari Spectral and Cloud Data

One of the primary challenges in satellite infrared (IR) quantitative precipitation estimates (QPEs) is accurately characterizing the nonlinear relationship between cloud properties and rainfall rates. This research proposes a deep neural network (DNN) method to classify clouds as rainy or non-rainy using brightness temperatures (BTs), reflectances (Refs), and cloud microphysical properties derived from the Advanced Himawari Imager (AHI) aboard the Himawari-8/9 satellite. The study incorporates cloud microphysical properties with BTs and Refs in the DNN model training process and conducts a comprehensive assessment of these features to elucidate their physical properties. The DNN-trained QPE models are validated by ground-based radar observation and compared to operational satellite-derived precipitation products like GSMaP and IMERG. The results indicate that including cloud microphysical properties enhances QPE model performance, with promising implications for real-time precipitation monitoring in East Asia.