Extraction Of Snow Cover In Xinjiang Based On FY-4A/AGRI

The average annual natural precipitation in Xinjiang is scarce.Snow melting water provides the main water resources for local river runoff and ecological development.Snow has ice/heat storage capacity which makes a great contribution to the stability of local ecosystem.Quantitative research with high precision on snow cover information in Xinjiang has an important impact on local residents' production and life development.With the development of meteorological remote sensing technology,remote sensing images with high spatial and temporal resolution also provide more possibilities for acquiring snow-related parameters and studying meteorological disaster warning.Compared with the traditional binary snow cover products,the estimation products based on the subcell scale of snow cover(FSC,Fractional Snow Cover)can provide more detailed and accurate snow cover area parameters for study areas with complex underlying surface types,severe mixed cells and blurred edge information.However,the spatial resolution of the data source greatly affects the accuracy of the remote sensing inversion model of snow cover.How to extract the information about snow cover accurately under the medium and low resolution image data is an important research task in this paper.In order to solve the above problems,we take FY-4A/AGRI data as the main remote sensing data source for inversion model of snow cover,makes full use of the space-time characteristics of AGRI data,fuses related geographic information data,uses deep learning method to carry out high-precision and quantitative estimation of snow cover in Xinjiang,and completes FSC proportional mapping;at the same time,it takes MOD10,the main international snow cover product.A1 was used as the source of contrast,and the high resolution image data provided by Landsat8 was used as the true value source to evaluate the accuracy of FSC inversion models based on different algorithms.The main contents and conclusions of this paper are as follows:(1)Resnet?FSC.Using the FY-4A/AGRI data of domestic stationary satellites as the main data source,the related geographic information data is used to fill the deficiency of spectral characterization ability;the high temporal resolution of AGRI data is used deeply to form cloud mask and reduce the probability of cloud and snow confusion;based on Deep Residual Network(Res Net)inverts snow cover in Xinjiang;This article uses MOD10?FSC method,BP-ANN?FSC method,spectrum??FSC method participate in comparison and verification of FSC method in snow cover inversion in Xinjiang;Landsat8-OLI high spatial resolution image data is used as the source of true value labels in this method model.The experimental results show that:Resnet?The mean of the determinant coefficients of the FSC method and the true value in the four sample areas is 0.46,compared with BP-ANN?The FSC method improves the prediction accuracy to a resolution of 500 m MOD10?FSC products have similar estimation accuracy.(2)Multiscale fusion Res?FSC.Because the Resnet?FSC method is affected by many factors,such as complex terrain in the study area and low resolution of remote sensing image data,it has great limitations in improving the estimation accuracy.To overcome these bottlenecks,Multiscale fusion Res-FSC method follows the main network structure of Resnet?FSC method and improves it on this basis.In the process of network feature extraction,the corresponding resolution image is integrated into the multi-scale information provided by different feature layers.By using the different expression advantages of deep and shallow features of deep network model,the network model's feature expression ability is enhanced and the prediction accuracy is improved.Finally,a high-precision 1000 m resolution snow cover estimation product is generated.The results of the four comparisons show that the average accuracy of EVS in the four sample areas is 0.03 higher than that of MOD10?FSC.From the final cartographic imaging results,it can be seen that the Multiscale fusion Res?FSC method is superior to the Resnet?FSC method before the network improvement in both estimation accuracy and spatial detail expression.