Research And Implementation Of Algorithms For Typical Land Covers Segmentation And Change Detection Of Remote Sensing Images In Arid Areas

Selecting the arid areas with fragile ecological environment to carry out the research of land cover segmentation and change monitoring is significant to meet the application demands of normalized monitoring of natural resources,spatial protection of western areas,and macro monitoring of important ecological barrier areas.The problems of uneven spatial distribution of desert oasis land covers,many small objects and blurred boundaries have become challenges for the research of fine extraction and change detection.Therefore,this thesis carries out the research of extraction and change detection of typical land covers of remote sensing images(RSIs)in arid areas based on the deep learning(DL)networks,which includes three aspects.(1)To address the limitations of traditional segmentation methods such as easy loss of detail information in the feature extraction stage and difficulty in obtaining global context information effectively,The segmentation model FPN＿PSA＿DLV3+ is proposed to improve the network performance in extracting multiscale features,integrating detail information and semantic information,and improving the ability of global and local spatial context information to capture finer edge and small object information in RSIs.Three land cover extraction datasets with more small objects and fine edges are constructed and make public based on Landsat 5/7/8 images for the two study areas.The FPN＿PSA＿DLV3+ method improves the segmentation performance of small objects and edges from 81.55% to 83.10% F1 and from 72.65% to 74.82% MIo U using only the RGB band.The FPN＿PSA＿DLV3+ network shows robust generalization across regions and sensors.(2)The existing change detection(CD)methods focus on extracting semantic features of deep changes and ignore fine-grained features,which are weak in capturing long-range spatio-temporal information and lead to leak classification for small change targets and edge smoothing of change feature types.We proposes a Pyramid-SCDFormer,a semantic change detection(SCD)model based on the Pyramid Transformer.The long spatial and temporal context information is captured,and the Shunted Self-Attention(SSA)module is introduced to subtly merge different semantic tokens in multi-headed self-attentive blocks to obtain multiscale change feature characteristics.In addition,a SCD dataset Landsat-SCD with more time-series images and change types is created and make public.Compared with existing advanced CD models,PyramidSCDFormer achieves 1.11/0.76%,0.57/0.50% and 8.75/8.59% MIo U/F1 improvement on the three datasets,respectively.For change types with change class proportions less than 1%,the model improves MIo U by7.17-19.53% on the Landsat-SCD dataset,strongly confirming the performance improvement of this thesis’ s model for small-scale change types and subtle edge recognition.(3)To address the problem that the research of DL networks for long time series change detection is still in depth,this thesis compares the FPN＿PSA＿DLV3+ change detection after classification and PyramidSCDFormer direct CD model on the test sets of Moso Bay reclamation and Tumushuk.The optimal CD model is selected to obtain the spatial distribution of four types of land covers,namely,farmland,building,water and desert,and analyze their spatial distribution and changes with the help of land use transfer matrix and dynamic attitude,and conduct change driving force analysis.The direct CD method is 26.5%/29.83%higher than the post-classification CD model in MIo U/F1 on the Moso Bay reclamation test set,and27.77%/28.31% higher in MIo U/F1 on the Tumushuk;the direct CD improves the MIo U of features with less than 1% of categories by a minimum of 13.85% and a maximum of 53.4% compared to the postclassification CD method.The land use dynamics of the two study areas are similar,with the land structure transformation of desert,farmland and building being more obvious,and the location of water being relatively fixed and changing smoothly.The overall land use transformation direction is from desert to farmland and from farmland to building.The study uses DL algorithms based on long time series satellite imagery to provide a feasible reference solution for monitoring the dynamics of land use and change.