Precipitation Inversion from MWHTS Data Using Tensorflow Framework

Abstract

The microwave humidity and temperature sounder (MWHTS) carried by FY-3D meteorological satellite is used to observe the vertical distribution of global atmospheric humidity, water vapor content, rainfall and other space meteorological data all day and all-weather, which plays an important role in atmospheric detection. In this paper, based on Tensorflow deep learning framework, using FY-3D satellite MWHTS data, we established precipitation inversion system from the perspective of neural network and random forest, and verified it with GPM rainfall product. The results show that both random forest (RFR) and neural network (MLP) model are feasible in precipitation inversion, The MSE is 1.76 and 1.84 respectively and the R2 is 0.79 and 0.78 respectively over ocean area. As the representative of ensemble learning, random forest has higher robustness and generalization ability. At the same time, it also displays that observation data of FY-3D MWHTS can show high application value in precipitation inversion.