Improving a Machine Learning Model for Satellite Precipitation Downscaling

Precipitation products at high space and time resolutions are critical for applications such as flood monitoring and water resources management. However, the space resolutions of commonly used satellite precipitation products, such as the NASA Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG, 10-km resolution) and NOAA/Climate Prediction Center morphing technique (CMORPH, 8-km resolution), are not sufficient for such applications. We have developed a generative adversarial network (GANs) based model, PrecipSRGAN, to create a super-resolution version of CMORPH, which has a spatial resolution of 4 km by 4 km. The super-resolution model uses the Stage IV quantitative precipitation estimates (QPEs) as references and takes advantage of the digital elevation model (DEM) to improve satellite-based precipitation feature extraction. Independent precipitation events were selected to evaluate the model performance. In addition, the deep learning-based downscaling performance was compared with a linear interpolation method. Based on the testing events and ground references, the PrecipSRGAN outperforms the linear interpolation-based satellite precipitation product given evaluation methods such as correlation (CC), normalized mean error (NME), etc. However, after comparing the deep learning-based downscaling product with the ground-truth dataset, it was noticed that the model inherits a terraindependent bias from the original CMORPH precipitation retrieval product that tends to overestimate precipitation in the Central Valley and underestimate precipitation over the mountainous regions, such as the Coast Ranges and Sierra Nevada. Therefore, this study attempts to improve the downscaling performance by adding weight factors to the loss function. The goal of such a weight matrix is to assign varying weights to the data points based on their precipitation levels. So that to give heavier penalty to the light rainy points and lighter penalty to the heavy rainy points. Testing results have demonstrated enhanced performance based on this adaption.