Study On Rural Settlement Extraction Method Based On Convolutional Neural Network And Spatiotemporal Variation Of Rural Settlement

With the acceleration of urbanization,China’s rural population is declining but the area of rural settlements is increasing,which leads to an increasingly imbalanced rural man-land relationship.For this reason,various regions actively promote the work of the rural-land comprehensive consolidation,which makes China’s rural settlements,especially in eastern inshore areas,are experiencing rapid changes.Therefore,real-time and rapid monitoring the number,distribution and spatiotemporal change characteristics of rural settlements is important for all these aspects,including improving the efficiency of rural land use,optimizing the spatial structure of rural settlements,and coordinating the sustainable development of urban and rural areas.The rapid development of aerospace technology and the launch of advanced satellite sensors have driven the field of remote sensing into the"big data"era.Coupled with the rise of deep learning,these aspects provide a new technical support for the rapid and accurate extraction of rural settlement.In this paper,Tongxiang City,located in the plain of northern Zhejiang Province,is taken as the study area.We construct the rural settlement extraction dataset separately by manually annotating high-resolution remote sensing images for two periods,2012and 2018.And based on the classic convolutional neural network Deep Lab v3+,the dual attention mechanism module is introduced in the coding area,and the decoding structure of layer-by-layer feature fusion is used in the decoding area.Finally,the rural settlement extraction model is constructed through the above optimizations.Through experimental comparison,it is proved that the optimized model can improve the accuracy of rural settlement extraction.Subsequently,the model was used to obtain information on Tongxiang’s rural settlements in 2012 and 2018.Then,we analyze the spatiotemporal variation and influencing factors of rural settlements from 2012 to 2018,in terms of area scale changes,spatial distribution and land use changes.The main results and conclusions of this paper are as follows.（1）Construction of rural settlement extraction datasets.We manually annotated two phases of remote sensing images with spatial resolution of 0.5m and 1m in Tongxiang City,annotating a total area of 304km~2,and performed data enhancement operations on the training set part.Finally,the datasets have 38,400 and 9,600 samples,respectively,which will be used for model training and validation.（2）Rural settlement extraction model construction based on improved Deep Lab v3+network.On the one hand,to address the deficiency of global information utilization prevalent in the original network,the dual-attention mechanism module is introduced in the network coding area to enable the model to capture long-range contextual information more effectively and improve the model feature extraction capability.On the other hand,to address the deficiency of the relatively simple structure of the decoding area of Deep Lab v3+network,the decoding method of layer-by-layer feature fusion is used to enhance the ability of the model to utilize shallow features and to enhance the recovery of detailed features.Finally,the construction of the rural settlement extraction model is completed based on the improvements in the above two aspects.（3）Model validation and extraction of Tongxiang’s rural settlements.We conduct a series of comparison experiments based on the constructed datasets.The results show that the dual attention mechanism module can improve the rural settlement extraction accuracy of the model,with the greatest improvement in the parallel structure.And the optimized decoding area structure of layer-by-layer feature fusion can enhance the spatial information recovery ability of the model and make the boundary extraction more accurate.In addition,in comparison with classical networks such as FCN-8s,U-Net,and Deep Lab v3+,the model in this paper achieves the best performance with98.23%,86.45%,75.54%,and 98.26%,85.59%,and 74.81%of PA,F1,and Io U on both datasets,respectively.This further demonstrates the effectiveness of the coding region optimization based on the introduction of a dual-attention mechanism module and the decoding region optimization method using layer-by-layer feature fusion in this paper.On this basis,we applied the model to extract information on Tongxiang’s rural settlements in 2012 and 2018.（4）Analysis of spatiotemporal variation and influencing factors of rural settlements.Based on the extracted information of Tongxiang’s rural settlements in2012 and 2018,we analyzed the spatiotemporal change characteristics in terms of area scale,spatial distribution and land use,and preliminarily analyzed their influencing factors.The results are as follows.The overall scale of Tongxiang’s rural settlements decreased from 2012 to 2018,and construction land moved from rural to urban areas.The distribution density presents regional differences,high in the southwest and low in the east.The new rural settlements are mainly in the west and north of the central city,which increase in dots,while the reduced rural settlements are mostly around the central area and the rest are more evenly distributed,most of which are scattered.The rural settlements are distributed in agglomeration,and the degree is increasing.Most of the reduced rural settlements have been reclaimed into farmland,and the dynamic balance of farmland has been achieved,while the fragmented farmland has been made more contiguous and scaled up through land remediation.Policy factors are the main factors affecting the spatial distribution and evolution of Tongxiang’s rural settlements,while location and population factors also have a greater impact on the distribution of rural settlements.