Desert Extraction Algorithms For Landsat 8 Images

Landsat 8 imagery has advantages such as short imaging period and large coverage,which is of great importance in land resources exploration,desertification monitoring and other fields.In desertification monitoring,desertification mainly occurs in boundary region of the desert region,so desert region extraction is one way for monitoring desertification.However,due to the complexity of features in desert,there are many features with similar spectral characteristics,which makes it a difficult problem to effectively extract desert from Landsat 8 images.Aiming at the above problems,this paper studies the desert extraction algorithm based on Landsat 8 images.The main work of this paper is as follows:(1)An improved multi-spectral desert superpixels generation algorithm based on entropy weight is presented.The edge fit of the existing desert extraction algorithms is not as good as expected.Compared with pixel-based algorithms,the superpixel-based algorithm has the characteristic of maintaining local spatial structure,and the detection results do not have pepper and salt effect..Therefore,this paper studies a desert extraction algorithm utilizing superpixels as the unit.However,the existing superpixel generation algorithm has the problem of inaccurate edge fitting in artificial building,transition zone and other non-desert land covers with similar spectral characteristics in the desert.In this paper,a band selection algorithm is proposed,in which bands are automatically grouped by improved correlation metric,and the entropy weight method is introduced to combine the desert distinguishable evaluation index and improved band evaluation index to select suitable bands for desert region extraction.Then,the band selection results are introduced into the existing Simple Liner Iterative Clustering(SLIC)algorithm,so that the input of the algorithm could be used to detect the effective bands of the desert.Moreover,the reliable pixel in the superpixel are used to calculate the features of the center point in the next iteration.Meanwhile,this paper presents an adaptive algorithm for superpixel compactness factor,which can automatically adjust the compactness factor using the spectral difference of various regions,to maintain the compactness and regularity of the shape of superpixels,and improve the edge fit of superpixels in complex terrain areas.The experimental data set uses images of the Taklamakan desert,Kumague desert and Kubuqi desert.The experiments show that the proposed algorithm has better performance than some existing algorithms,in which the edge boundary recall is better about 0.14,the under-segmentation error is better about 0.02,and about 0.16 is the achievable segmentation accuracy.It is verified that the proposed algorithm has advantages over the existing algorithms in edge fit of artificial buildings and desert transition region.(2)A desert extraction algorithm for Landsat 8 image is presented.Due to the problem that the existing desert extraction algorithm cannot effectively remove the artificial buildings and other non-desert land covers that have similar spectral characteristics in the desert.In this paper,a rough extraction algorithm based on random walk and superpixels is proposed to extract desert regions,in order to remove non-desert features.Then this paper presents a regular spatial structure to remove cultivated land.For the desert transition zone and the Gobi,the Normalized difference Desert Index(NDDI)is provided to judge foreground by pixel after removing the background objects with regular structure.Then the transition zone and the Gobi and other irregular ground objects around the desert could be removed,so as to improve the precision of desert extraction.The experimental data set are obtained from the Kumtag desert,Taklamakan desert,and Kubuqi desert.Experiments show that the overall accuracy could be better about 3%and the Kappa coefficient increased by about 0.06 in comparison with some existing algorithms,which demonstrates that the algorithm in this paper has superority over the existing algorithms in large size image and non-desert ground object removal.