CNN-LSTM Networks Based Sand and Dust Storms Monitoring Model Using FY-4A Satellite Data
Zhao Zhen, Zihang Li, Fei Wang, Fei Xu, Guoqing Li, Hongjun Zhao, Hui Ma, Yiran Zhang, Xinxin Ge, Jianan Li
Abstract
Improving the accuracy of sand and dust storms monitoring can provide effective support for the safety warning of extreme weather for the power system, which in turn can enhance the power system's emergency supply capacity as well as safeguard the normal production and life of human beings. In this paper, a hybrid sand and dust monitoring model based on one-dimensional convolutional neural network (CNN) and long short-term memory (LSTM) network is proposed. The normalized dust index (NDDI), CNN model and 1DCNN-LSTM hybrid model are utilized in conjunction with Meteosat IV (FY-4A). Sand and dust storms in the Taklamakan Desert in southern Xinjiang were monitored and studied using channel scanning imaging radiometer AGRI (Advanced Geostationary Radiometer) data. The results show that the NDDI sand and dust indices determined from images at different times require the use of different thresholds to identify the sand and dust zones. Recognition errors exist in both covered and desert areas. According to several sand and dust storms event tests, the monitoring model based on the 1DCNN-LSTM network can achieve 91.42% recognition accuracy, which is a stronger monitoring capability compared to the CNN model as well as traditional models. In practical applications, the 1DCNN-LSTM model outperforms the CNN model in dealing with sand and dust and non-sand boundaries. In addition, the 1DCNN-LSTM model can recognize dust storms more accurately under a small amount of cloud occlusion.