Research On Cloud Detection Method Of Landsat 8 Images Based On Neural Network

In recent years,with the rise of remote sensing technology,the demand for satellite data in all walks of life is also increasing.As one of the most important satellite data sources for earth observation,Landsat series data has gradually become one of the most effective data for observing surface features and studying the application of earth system with the advantages of long-term continuity,global coverage and appropriate spatial-temporal resolution.However,cloud pollution has a serious impact on Landsat series images,and cloud also has a significant impact on the inversion of surface parameters and atmospheric parameters,which greatly reduces the utilization of remote sensing images.Therefore,cloud detection is an indispensable step in remote sensing image processing.The existing cloud detection methods often rely on specific sensors and bands,and require strict parameters.Based on the theory of deep learning,this paper studies the cloud detection method of Landsat 8 data based on neural network.The research mainly focuses on the following aspects: first,it summarizes the research status of cloud recognition at home and abroad in detail,and classifies and summarizes the methods.This paper summarizes the features of cloud in remote sensing image,studies and implements the most popular algorithm of Fmark cloud recognition,and introduces the concept of neural network according to the deficiency of its easily lost spatial features.Secondly,according to the actual needs of cloud detection,three kinds of neural networks for cloud detection are designed,which are CNN,U-net and SCU-net.The real data set is used to train it,and the cloud recognition model with the best parameters is selected according to the training process.Then,according to the spatial and spectral characteristics of the cloud,a hybrid model of cloud distortion is proposed,and the cloud data with real spectral characteristics is simulated by the hybrid model.The simulation data,SPARCS and L8 Biome are used to verify the designed network,and compared with the traditional methods,qualitative and quantitative analysis are carried out respectively.Finally,according to the images with different underlying surfaces and different cloud shapes,the above methods are used to test,analyze the difference of the results of processing cloud data in different situations,and get the applicable regularity of the method.The experimental results show that this algorithm can effectively extract the spatial features of the image,and multi-scale fusion of spatial and spectral information,so that the accuracy of cloud detection is improved.