Research On Cloud Detection And Precipitation Extrapolation Forecast In Tibet Based On Fengyun-4 Satellite Images

Cloud and precipitation are important indicators to represent weather and climate change,and are important components of global energy cycle.Using joint observations of multiple satellite instruments to study clouds and precipitations has become a hot topic in the field of remote sensing and climate change.As the main area of the Qinghai-Tibet Plateau,Tibet can affect the weather and climate over regions through its mechanical and thermodynamic effects.Thus,it is of great significance to study the cloud and precipitation on the plateau.Meanwhile,due to the complex terrain conditions and uneven distribution of precipitation seasons on Tibet.It is vulnerable to severe convective weather,such as precipitation and gale in spring and summer,and resulting in various geological disasters,which seriously threaten the life safety of local residents and the normal operation of highway traffic.Therefore,this paper carried out a large-scale cloud detection and precipitation forecast work in Tibet.In view of the rare distribution of meteorological observation stations and radar base stations in this area,it is impossible to carry out large-scale weather monitoring effectively.In this paper,Feng Yun-4A(FY-4A),a high-time resolution and multi-observation band satellite,is used as the data source.Under the background of remote sensing big data,the deep learning method is used to fully explored the corresponding relationship between satellite multi-band data and cloud cluster and precipitation,so as to automatically carry out cloud detection,precipitation monitoring and forecasting on Tibet.Firstly,this paper takes the following as the research object.Due to the fact that some small clouds only occupy a few pixels in geostationary satellite images with low spatial resolution,the existing network has poor segmentation ability for these small targets.To tackle this problem,a new neural network named U-High Resolution Network(U-HRNet)which can produce high-resolution representation is established.By connecting the convolution structure from high resolution to low resolution in parallel and fusing multi-scale features for many times,the loss of detail information is reduced,and accurate segmentation is achieved.The Mean Intersection over Union(MIou),PA(Pixel Accuracy)and F1-scores of detection results are 94.03%,94.21% and 94.11% respectively.The accuracy of the method is evaluated by manually labelled ground truth against different methods using objective evaluation indices.The results proved the proposed U-HRNet performs well on FY-4A's images and can effectively detect incorrect areas of the cloud mask products of the National Satellite Meteorological Centre(NSMC),depicting outperformance over existing methods.Secondly,in view of the characteristics of precipitation cloud clusters and the problem of distinguishing precipitation clouds from non-precipitation clouds,the network structure based on the U-HRNet with high-resolution representations is improved and the attention mechanism module is introduced.The new network structure is applied to the small sample data set of precipitation cloud detection and precipitation intensity grade estimation.Taking GPM(Global prediction measurement)precipitation inversion product as the ground truth and integrating DEM data into the network input,the corresponding relationship between FY-4A multi-observation band and precipitation is established,so as to realize the precipitation monitoring on Tibet.The values of Probability of Detection(POD)and Critical Success Index(CSI)in precipitation intensity estimation results are 0.918 and 0.916,which are the highest among several comparison methods,and the False Alarm Ratio(FAR)value is the lowest of0.045.Then,taking FY-4A time series data as input,spatiotemporal prediction network is used to predict satellite images and multi-channel data.Aiming at the problems of unsatisfactory accuracy of the existing method extrapolation results,blurred images and poor data quality.Use the deformable convolution structure to capture the information of deformation target for improve the network's extraction of cloud features.Combined with the idea of optical flow,cloud detection results and the network structure of Generative Adversarial Network(GAN),the joint loss function is constructed.This paper proposes a adversarial cloud image prediction model named Deformable Convolution GRU2Pixel(DCG2Pix)for achieve high-precision extrapolation of satellite cloud images and multi-channel data after 30 minutes.By comparing with actual observations,the accuracy of the predicted data is verified.Finally,the predicted data is input into the established precipitation intensity level estimation model to realize the precipitation nowcast over Tibet.