Research On Rural Settlement Extraction Method Of Multi-source Remote Sensing Based On Machine Learning

Rural settlements—that is,settlement areas for rural residents to produce and live in—are important carrier of rural development.In recent years,due to China’s rapid industrialization and continuous socioeconomic progress,the spatial distribution of rural settlements has undergone significant transformation,leading to negative effects such as“hollow villages”and“unbalanced development of the human-land relationship”.China is a large agricultural country.The disorderly development of rural settlements encroached on a large amount of fertile land,which is not conducive to the intensive use of land resources and affects the modernization process of the countryside.Therefore,a comprehensive understanding of the spatial distribution and structural scale of rural settlements is an important guarantee for the scientific management,sustainable development of rural settlements and the rational development and utilization of rural land resources in China.Remote sensing technology has become an important means of accurately extracting land resource information due to its large-scale monitoring and high timeliness.However,there are still many difficulties in accurately and efficiently extracting rural settlements.Therefore,based on the Google Earth Engine(GEE)cloud computing platform,this paper first designed a machine learning automated algorithm model for extracting rural settlements using Landsat images and Sentinel images,supplemented by nightlight data and DEM data.Then,based on the Tensorflow deep learning framework,this paper designed a deep learning model for automatic extraction of rural settlements using Sentinel-2 multispectral image data.The main research results of this paper include the following three aspects:(1)Research on automatic extraction of rural settlements based on Landsat images and CART algorithmThrough Google Earth high-resolution remote sensing images and literature surveys,we found that rural settlements in Xuzhou mainly include low-density large-agglomerate types and medium-density broadband types.This paper integrated Landsat 8 OLI remote sensing images,VIIRS-DNB night light data,SRTM DEM data to generate spectrum,texture and terrain multi-dimensional classification feature set.Then we integrated these features to build CART(Classification and Regression Tree)model,and finally realized accurate and automatic extraction of rural settlements in Xuzhou.This method got a high extraction accuracy,with an overall classification accuracy of 90.2%and a Kappa coefficient of 0.865.The analysis results showed that the area of rural settlements in Xuzhou is about 1545.02 km~2,accounting for 13.8%of the total area.Rural settlements around Feng County,northwest of Pei County,northern Pizhou,northern Xinyi and southwest of Jiawang District are mainly low-density and large-mass distributed;Rural settlements near Xuzhou City,near Suining County and southern Xinyi City are mainly medium-density and broadband distributed.(2)Research on automatic extraction of rural settlements based on multi-source Sentinel data and Random Forest algorithmThe Sentinel-1/2 data were used to improve the automated extraction algorithm of rural settlements.Compared with the Landsat extraction model,this model had higher extraction efficiency and stronger universality.In detail,this model introduced Sentinel-1 SAR data with rich backscatter coefficients and spatial texture information,which had effectively improved the accuracy of land resource extraction.Meanwhile,the partition strategy and the idea of migration were used to greatly improve the extraction effect.Accordingly,the extraction of the distribution information of rural settlements in the Yangtze River Delta in 2019 was finally realized.The overall accuracy and Kappa coefficient were 96%and 0.84,respectively.The final mapping results provided a detailed spatial distribution of the rural settlements in the Yangtze River Delta and revealed that the total area of rural settlements is approximately 32121.1 km~2,accounting for 17.41%of the total area.The high-density rural settlements are mainly distributed in the Northern Plain and East Coast,while the low-density rural settlements are located in the Central Hills and Southern Mountain.(3)Research on automatic extraction of rural settlements based on Sentinel-2multispectral data and fully convolutional neural network modelBased on Sentinel-2 optical data,we used the full convolutional neural network to train a model suitable for the extraction of rural settlements in the Yangtze River Delta to extract rural settlements in this region,and then compared it with the extraction based on shallow machine learning algorithms.The results showed that the overall accuracy and Kappa coefficient of rural settlements in the Yangtze River Delta extracted by the fully convolutional neural network model have reached 98%and 0.91,respectively,which has been greatly improved compared with the random forest algorithm.The fully convolutional neural network algorithm obtained better results for large-scale blocky rural settlements than the extraction results based on the random forest model.However,for some small-scale scattered rural residential areas,they were slightly worse.There was a certain degree of confusion in the extraction results of the fully convolutional neural network algorithm,which may lead to a certain overestimation of small-scale scattered rural residential areas.The research results of this paper can provide important data and technical support for rural development planning and rural environmental assessment in the Yangtze River Delta.The automatic extraction algorithm of multi-source remote sensing data proposed in this paper also has certain reference value for the extraction of other ground feature information at the national and even global scales.