资源环境科技发展态势分析平台

Real-time precipitation estimation from satellite observations is still a challenging issue. In this paper, we propose an effective and efficient model based on convolutional neural network (CNN) to estimate precipitation over Xinjiang, the driest region in China, from Fengyun-2 satellite data. Our network mainly consists of two modules: the precipitation identification module and the precipitation estimation module. The first module aims at identifying the given region as precipitation or non-precipitation, while the second one focuses on estimating the specific precipitation amount of the identified region. In addition, considering that topography generally have effects on the quantitative precipitation estimation (QPE) task, we therefore incorporate them into the second module. To evaluate the effectiveness of our proposed model over the national stations, we compare it with not only the Integrated Multi-satellite Retrievals for the Global Precipitation Measurement (GPM) mission (IMERG), but also another two near-real-time products named Global Satellite Mapping of Precipitation Near-Real-Time product (GSMaP NRT) and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Cloud Classification System (PERSIANN-CCS). Be- sides, two precipitation events are selected to validate the generalization ability of our proposed model in Xinjiang, and the verification experiment is conducted upon the precipitation product IMERG Final run. Experimental results show that, in comparison with the rain gauges records, the precipitation identification accuracy and correlation coef- ficient of our proposed model are 0.93 and 0.21, while the precipitation estimation RMSE and MSE of our proposed model are 0.25 and 0.038 mm/h, respectively. For the two precipitation events, our proposed model demonstrates high consistency with IMERG Final run in precipitation identification, but it shows a tendency to underestimate the heavy rainfall estimated by IMERG Final run. More importantly, the implementation time of our proposed model on the hourly testing dataset is only 5.5 s. Therefore, our proposed model can achieve near-real-time results, which is valuable for precipitation nowcasting applications.