| Xinjiang is one of the three stable snow cover regions in China,which is rich in seasonal snow resources.Its snowmelt runoff is also the source of many important rivers.The healthy development of animal husbandry has a profound impact on the prosperity and stability of the region.Snow disaster in pastoral areas is the most frequent and influential one in China,so it is very important to monitor the snow in Xinjiang with large scale and high frequency by remote sensing.In the existing snow monitoring methods,the combination of ground station observation data and remote sensing satellite detection is often used.However,Xinjiang is vast and sparsely populated,and there are not enough ground monitoring stations.Most of the products are produced abroad,which is not conducive to the promotion and application of remote sensing snow monitoring in China.Therefore,this paper takes the new generation of domestic FY-4 satellite as the main data source,uses its ultra-high time resolution characteristics,integrates various geographic information data that affect snow recognition,and uses deep learning method to adaptively extract snow feature information to generate hourly snow.Then,on the basis of the generated snow discrimination results,the multi-temporal cloud filtering method is mainly adopted,supplemented by the combination of meteorological and altitude cloud filtering method to achieve cloud removal,so as to generate the daily snow-less products in Xinjiang and effectively improve its application value.The main research contents and conclusions are as follows:(1)Snow cover recognition is based on FY-4A/AGRI 2km resolution.To solve the problem that existing snow cover products are susceptible to complex topographic landforms,underlying surface types and cloud cover,which leads to the reduced accuracy of snow cover identification,a method of snow cover identification based on multi-feature time series fusion of the AGRI data and geographic information data of Feng Yun 4 star A is presented by using deep learning method: multitemporal FY-4A/AGRI multispectral remote sensing data,which can be used to identify snow cover in the future.Taking the multitemporal FY-4A/AGRI multispectral remote sensing data,elevation,aspect,slope,land cover type and other topographic information as the model input,and the high spatial resolution snow cover map extracted by Landsat8-OLI as the "true value" label,the snow recognition model based on convolution neural network is constructed and trained,so as to effectively distinguish the complex terrain and the underlying surface area of Xinjiang.Finally,the product of hourly snow coverage is obtained.The accuracy of this method is higher than that of MOD10A1 and MYD10A1,which are the main international MODIS snow products.The accuracy of snow classification in snow season is 94.15%,which significantly reduces the misclassification rate of cloud and snow.(2)Research on snow products and multi-stage intraday cloud removal method based on deep learning.According to the FY-4A remote sensing snow products mentioned above,the method of multi-temporal time cloud filtering is mainly adopted to generate preliminary daily snow-less products according to the adjusted window period.By using the characteristics of AGRI hyperspectral,a cloud motion index is established to identify the rain and snow which are difficult to distinguish in the image.Then,an elevation filter algorithm based on CLDAS meteorological conditions is used to remove the cloud.Combined with the SNOWL snowline algorithm,a certain range of snow maps are identified.Finally,combined with the three-stage recognition results,a multi-stage fusion of less cloud products is established to reduce the cloud coverage of snow recognition products in Xinjiang.The final average annual cloud cover of FY-4A/CL less cloud products decreases from 46.98% to 31.65%,and the average annual classification accuracy still reaches 93.16%.Experiments show that this method has great advantages in cloud removal ability and reliability. |
PDF Full Text Request