Research On Cloud Detection And Cloud Image Prediction Method Based On FY-4A Satellite

Over 66% of the earth's surface is often covered by clouds,and the change of clouds indicate the occurrence of various weather phenomena.When the rapid changes of clouds occur in catastrophic weather such as rainstorm,thunder and lightning,they will have a great impact on human life,and even threaten the safety of human life.Especially in Tibet,the strong convective weather in summer can easily lead to natural disasters.Therefore,it is necessary to monitor the changes of cloud cluster in the area to prevent disasters.FY-4A has realized the real-time observation of clouds,which can provide high temporal resolution and multi-channel observation data.With the help of this data to analyze the dynamic changes of clouds,more accurate weather forecasts can be achieved.Therefore,Therefore,this paper takes FY-4A as the data source and combines deep learning technology to conduct cloud detection and cloud image prediction research in Tibet.To solve the problem that existing deep learning models have large volume and weak ability to capture cloud features,a lightweight cloud detection model based on an improved UNet is proposed.The model takes U-Net as the basic framework and uses all-weather cloud image as input data.The residual module and the convolutional block attention module are integrated into U-Net to improve the model's ability to extract cloud features,and hardly increase the computational complexity of the model.In addition,the use of depthwise separable convolution in the model can generate a lightweight cloud detection model without degrading the model performance,providing the possibility of embedding the model into mobile devices.Experimental results show that the cloud detection model proposed in this paper can detect a large number of broken clouds and thin clouds,and the m Io U reaches 92.21%,which is 1.79%higher than that of U-Net.In order to solve the problems of fuzzy and low precision of cloud image generated by the existing time series prediction models,a cloud image prediction model based on 3D generative adversarial network was proposed.A cloud image prediction model based on 3D generative adversarial network is proposed.The model uses historical observations of the FY-4A to predict future satellite image and infrared channel brightness temperature.Three-dimensional convolution is applied in the model,and the training method of generative adversarial is adopted to improve the ability of extracting spatio-temporal features of the model.In addition,the definition of the predicted image and the authenticity of the infrared brightness temperature are further improved by improving the loss function.Through experiments,it is verified that the cloud image prediction method proposed in this paper can generate clearer image and accurate infrared brightness temperature,and the SSIM index of image reaches 0.85,and the RMSE index of infrared brightness temperature is less than 10 K.In order to analyze the availability of the forecast data,the improved U-Net is used to perform cloud detection on the forecast cloud image.The results show that the dynamic changes of the forecast cloud image are similar to the actual cloud image,which indicates that the proposed cloud image prediction method can provide reliable data support for the weather forecast.