Sand and Dust Storms Monitoring Using FY-4A Satellite Data based on Convolutional LSTM Networks

Abstract

Improving the sand and dust storms monitoring capability can provide early warning information for sandstorms, and effectively improve the PV power prediction accuracy under sand and dust storms weather. In this paper, a hybrid model dust monitoring method based on a one-dimensional convolutional neural network (CNN) and a long short-term memory (LSTM) network is proposed. Using the normalized difference dust index (NDDI), CNN model, and 1DCNN-LSTM hybrid model, combined with the number four meteorological satellite (FY-4A). The channel scanning imaging radiometer AGRI (Advanced Geostationary Radiation Imager) data is used to monitor and study the sand and dust storms of the Taklimakan Desert in southern Xinjiang. The results show that the NDDI dust index established by images at different times needs to take different thresholds to identify the dust area. There are misidentifications in the coverage area as well as in the desert area. The sand and dust storms monitoring model established based on CNN network and 1DCNN-LSTM network, the accuracy (Accuracy) and loss function (Loss) of training samples and test samples are 99.9% and 1%, respectively, which has strong sand and dust storms monitoring capabilities. In practical applications, the 1DCNN-LSTM model is better than the CNN model in processing the boundary between sand and non-sand dust. In addition, the 1DCNN-LSTM model can also more accurately identify sand and dust storms in the case of a small amount of cloud occlusion area.