Convolutional Neural Network for Rain/No-Rain Detection from Satellite Observations

Satellites are playing an important role in monitoring the global precipitation due to their full disc coverage and high spatiotemporal resolution. A lot of statistical algorithms have been presented in describing the relationships between satellite observations and precipitation. However, most of these algorithms utilize artificially extracted features, like the mean and variance of brightness temperature in a square area, which can not fully mine the data information. In this paper, we use a convolutional neural network (CNN) based method, which has been proved to be powerful in extracting features automatically, to delineate rain/no-rain (R/NR) areas from observations of the next-generation geostationary meteorological satellite Himawari-8. Observations from about 10000 rain gauges covering a period of half a month in southeastern China were used to train and test the method. To provide accurate simulations separately on rainy and dry seasons, the dataset was split into different subsets according to the amount of precipitation. The commonly used self-organizing feature map (SOFM) method was used as a reference in the experiment. Our results show that, the CNN method performs better than the SOFM method, with ETS score increases ranging from 0.05 to 0.1 under different subsets, thus provides a new way for R/NR detection.