Monitoring Of Land Desertification Based On Multi-Scale Segmentation From Remote Sensing Date

Desertification is a serious ecological environment problem in the world.Effective monitoring of desertified land is an important prerequisite for its protection and governance.Traditional pixel-based classification methods often have the problems of spectral confusion as well as salt and pepper phenomenon in desertification information extraction.In this study,multi-source remote sensing images and land cover products were selected as the data source to determine the optimal scale of image segmentation based on multi-scale segmentation technology.Considering the features of image spectral,geometry and relevant thematic indices,the object-oriented random forest classification method was adopted to conduct remote sensing monitoring of land desertification in Mu Us Sandy Land from 2001 to 2021by using long time series Landsat data.The spatial distribution and migration of desertification in different periods were analyzed in order to provide decision-making suggestions for desertification prevention and control.The main contents and conclusions are as follows:(1)Aiming at the characteristics of desertification land in the study area,this study used mean variance method and maximum area method from the perspective of spectrum and space,and combined artificial visual observation to determine the optimal scale of desertification information extraction.It was found that the optimal segmentation scale for Sentinel 10m resolution image was 460,Sentinel 20m resolution image was 240,Landsat30m resolution image was 140,Sentinel 60m resolution image was 70,and MODIS 500m resolution image was 27.The results indicated that when extracting information from a certain type of ground object,the optimal segmentation scale gradually decreased as the spatial resolution of the image decreased.(2)Taking Landsat image as an example,different classifiers were used to classify the land desertification degree in the study area based on the optimal scale.It was found that the random forest classification method was more suitable for remote sensing monitoring of land desertification in Mu Us region.Based on the optimal segmentation scale,the random forest classifier was applied to extract desertification information from multi-source data.The research found that the extraction result of 20m resolution data was higher than 30m,10m,60m and 500m.The results indicated that high spatial resolution of images did not necessarily mean high extraction accuracy,and low resolution data was not suitable for object-oriented remote sensing monitoring of land desertification.(3)Compared with the classification results based on pixel,the object-oriented classification method based on the optimal segmentation used in this study comprehensively considered the spectral,geometric and thematic index characteristics of segmented objects,and used the spatial feature optimization tool to select the features,which effectively overcome the spectral confusion as well as salt-and-pepper phenomenon.This method has achieved good results in the extraction accuracy of desertification land in different degrees,with the overall accuracy increased by 5.616%and Kappa coefficient increased by 0.0774,which can be better applied to the extraction of desertification information in a large range.(4)The remote sensing monitoring of land desertification was conducted in Mu Us region from 2001 to 2021.The results showed that the desertification land area in the study area increased first and then decreased in the past 20 years,with a total decrease of587.12km~2.The coverage of moderate desertification land was relatively extensive,accounting for more than 37%.By calculating the transition matrix and gravity center coordinates,the change of desertification land of different degrees in different periods was systematically analyzed.The results showed that the gravity center of severe desertification shifted westward during 2016-2021,and the gravity center of desertification in other periods was concentrated without obvious expansion phenomenon,and the control effect was good.In terms of spatial distribution,the effect of desertification control was better in the eastern and southeastern regions,and the areas with non-desertification and mild desertification have increased significantly,while the western region was prone to recurrent desertification,and there were still some areas with moderate and severe desertification that need to be strengthened.In the future governance process,reasonable prevention and control should be considered in multiple aspects.