Research And Application Of Snow Recognition Algorithm Using High Temporal-spatial Resolution Remote Sensing Imageries

Due to the unique geographical topography and climate characteristics of the Qinghai-Tibet Plateau,it has brought great challenges to the inversion of snow cover distribution by optical remote sensing data.First of all,the altitude of the Qinghai-Tibet Plateau(TP)is relatively high,the elevation difference between the east and west of the TP is large,the range of the QinghaiTibet Plateau is wide,and the lighting conditions are different.Secondly,due to the cloudy winter in the Qinghai-Tibet Plateau,it is difficult for optical sensors using polar-orbiting satellites as carriers to detect the snow under the clouds.Due to the characteristics of the observation data in the same area,the geostationary satellites can complete the objects under the clouds through the time difference.Therefore,the optical remote sensing image of the multi-channel scanning imaging radiometer of China's new-generation geostationary satellite Fengyun-4A was used as the data source,and a snow cover extraction algorithm combining snow optical characteristics,texture features and motion features was proposed for the TP.The research work of this paper is mainly divided into the following four parts:First,we introduce the used satellite imageries,ice and snow products and ground weather stations' data information,and put forward the pre-processing process and algorithm for FY-4A / AGRI,Landsat-8 OLI/TRIS,and comparative data.These data lay the foundation for later research work.Then,an unsupervised algorithm is proposed to overcome the difficulty in manually labeling remote sensing imageries.This algorithm extract snow cover based on Gaussian mixture model.It integrates the threshold method and Gaussian mixture model clustering algorithm to achieve these goals.First of all,it extracts the features of snow based on the normalized differential snow index,and set threshold for the relevant band.Then,the means and covariance of the extracted "pseudo-label" are calculated for initializing the Gaussian mixture model.Finally,the EM iterative algorithm is used to update model parameters iteratively,and the results are output after the iteration conditions are met.The experimental results show that without human intervention,the recognition result of the algorithm can reach 77.09%,which is higher than other international mainstream ice and snow products,1.3% higher than MODIS snow products,and the snow detection rate has increased by 37.72%.Compared with microwave data fusion,the accuracy of ice and snow products is improved by up to 3.4%,which proves that the algorithm is reasonable,feasible,and effective.It reduces the difficulty of manual labeling without losing accuracy,and the results can be used as a reliable data source for later snow monitoring research.In order to improve the spatial resolution of the results and reduce the interference caused by the mixed pixels,we use the FY-4A / AGRI higher spatial resolution remote sensing data to complete the snow extraction.Because clouds and snow have similar optical characteristics in visible light and near-infrared bands that contained in this data,a three-dimensional orientation of gradient algorithm based on snow' optical features,texture features,and motion features is proposed.It detects the edge of imageries in the time dimension,and uses an edge detector to extract the edge of clouds and snow in three dimensions.Finally,the original optical image data,spatial gradient and 3-dimension detection results are used as input,and a linear support vector machine is used to separate the snow from clouds.Experimental results show that the algorithm has the highest accuracy,reaching 78.30%,which is 1.21% higher than the previous algorithm.The error detection rate ranks second among similar snow products,only 9.75%.A snowfall tracking process for the Shigaze also conforms to previous weather forecast and precipitation information.The results show that the algorithm is reasonable,feasible,and reliable,and it can achieve an approximate or even better extraction accuracy while maintaining fewer band characteristics.Finally,multi-temporal remote sensing images of FY-4A satellite data were used to monitor and track a strong snowfall process on the Qinghai-Tibet Plateau.In order to highlight the advantages of high-temporal data of FY-4A / AGRI,the change process from daily snow accumulation was demonstrated.From the daily snow change process to the hourly snow change process.Due to the large amount of data and the time-intensive observation data,it is possible to obtain the spatiotemporal change characteristics of snow cover within an hour,a day,or multiple days,which is of great significance for the use of snow cover resources on the TP.