Study On Deep Learning-based Extraction Of Rural Housing And Its Spatiotemporal Evolution

Rural settlements are the carriers of rural production and life,and rural housing is the main element of rural residential land.The spatial clustering results of rural housing data can be used to characterize the range of rural settlements,to understand the dynamic changes of rural settlements,which provides a scientific basis for the optimization of rural settlement layout,the intensive and economical use of land,rural spatial reconstruction,and promotion of rural revitalization.The increasing abundance of remote sensing data products with high spatial resolution makes the accurate identification and rapid mapping of rural buildings possible.With the rapid development of machine learning,automatic interpretation of high-resolution remote sensing images based on deep learning technology has become a research hotspot in the remote sensing community.This study takes a typical rural area in the water network plain of Yangtze River Delta as the study area and uses multi-source high-resolution remote sensing data to study the extraction method of rural housing information based on deep learning,and applies the results of rural housing spatial clustering to represent rural settlements.Finally,an in-depth discussion on the spatial and temporal dynamics of the spatial distribution characteristics of rural residential areas is conducted.The main contents and conclusions of this paper are as follows:(1)Based on high spatial resolution aerial images and full convolutional networks(FCNs),an automatic building extraction method(RFA-UNet)applicable to sub-meter spatial resolution images is proposed.To address the problem of feature redundancy and representation divergence caused by semantic inconsistency when fusing features from different levels,an attention-based re-weighting technique is introduced into the skip connection structure of the network.The learnable attention weight is used to adjust the response of shallow features in channel and spatial dimensions,to enhance the consistency of feature representation.The results obtained from the experiments conducted on three aerial imagery building datasets show that joint attention module(RFA)can effectively improve the recognition of classical segmentation models such as U-Net.Compared with other methods,the results of RFA-UNet are completer and more accurate,which can meet the demand of high accuracy extraction of buildings in high resolution images.(2)Based on high spatial resolution(HSR)satellite images and FCNs,a multiscale dilated residual network for automatic rural housing extraction is proposed.The rural man-made structures in HSR images have irregular morphologies and complex distribution patterns,thus the accuracy of traditional classification methods is relatively low.Considering the characteristics of spectral differences and large spatial scale variations within rural housing classes,a dilated residual convolutional network(Dilated-Res Net)is used to extract high resolution feature maps to obtain more spatial context information,followed by the optimization of features using multi-scale feature fusion structure and channel attention module,which improves the ability of the network to classify rural housing.The results of the experiments conducted on the Gaofen-2 images show that the proposed approach can effectively distinguish rural housing from other man-made structures and obtain more accurate rural housing classification results,and the F1-socre of the two types of rural housing is better than85%.In practical terms,the framework proposed in this study can be extended to a wider range of rural areas,providing an important reference for large-scale cadastral surveys or rural housing change monitoring at the regional scale.(3)Based on semi-supervised learning and ensemble learning,a rural housing extraction method suitable for multi-source multi-temporal high-resolution images is proposed,and the rural housing information of Tongxiang in 2005,2012 and 2018 is successfully extracted.Current remote sensing classification methods based on deep learning rely on a large number of labeled samples,and the high cost of manual labeling makes it difficult for such supervised models to cope with complex scenarios constructed from multi-source and multi-temporal remote sensing images.To solve the above problems,the study proposes a deep ensemble network framework that integrates multiple semantic segmentation models built on few samples,and it generates pseudolabels by filtering the ensemble prediction with high confidence from unlabeled images for subsequent iterative optimization.The experimental results show that the overall accuracy of the three years is better than 83%,and the improvement of the average accuracy is 1.6%,3.1% and 4.0%,respectively,which proves that the semi-supervised ensemble learning strategy can effectively alleviate the problem of insufficient labeled samples and enhance the performance of the deep model.Moreover,the proposed semisupervised ensembled learning method can provide reference for the intelligent interpretation of remote sensing images and improve the utilization of massive remote sensing images.(4)Based on the results of rural housing extraction in Tongxiang in 2005,2012 and 2018,the rural settlement of Tongxiang in 2005,2012 and 2018 is quantified by using a percolation-based clustering algorithm(CCA),and exploratory spatial data analysis method is used to study the spatial-temporal dynamic pattern of rural settlements.The results show that: 1)the percolation-based CCA method can perform spatial clustering according to the building density attributes of rural housing grids,and achieve the approximate extraction of rural settlement extent at the grid scale;2)The total area of rural settlements in Tongxiang increased and then decreased from 2005 to2018,with the total area decreasing by 14.5%;the spatial distribution of rural settlements in Tongxiang during 2005-2018 is spatially clustered,and is dominated by low-low agglomeration and low-high agglomeration;3)from 2005 to 2012,the area of rural settlements increased,and the two hot spots of settlement growth are the southwest region and the north region of Tongxiang,which are far away from the downtown area.From 2012 to 2018,the area of rural settlements decreased significantly,and the central urban area and its surrounding towns are the high-value agglomeration areas of settlement area reduction.