Rural Built-Up Area Extraction From Large-scale Remote Sensing Images Using Spectral Residual Methods With Embedded Deep Neural Network

The identification and observation of built-up areas can be used to understand the land changes and make reference to the promotion of economic development.In vigorous development of rural construction,the identification and extraction of the rural built-up area information is particularly important.By quickly and accurately obtaining rural built-up area information and studying the identified rural built-up area,we can provide scientific basis for scientific planning of rural development and construction.As one of the most important targets in remote sensing images,a large number of researchers have studied the extraction of built-up area in remote sensing images,but most of them are on urban built-up areas and buildings,while the research in this paper mainly focuses on rural built-up area and conducts experimental investigation on the extraction of rural built-up area in a large-scale area.The eastern foothills of Taihang Mountains are selected as the study area,and Gaofen-1(GF-1),Resource-3(ZY-3)and World View-2(WV-2)are used as the experimental data sources for the study.The rapid development of deep learning networks in recent years has brought opportunities for extracting built-up areas on large scale remote sensing images.In this paper,we embedded the Spectral Residual(SR)method into a deep neural network for rural built-up areas extraction,which consists of two main processes:rough localization and accurate extraction.Firstly,the Faster R-CNN(Regions with Convolutional Neural Network)model in the deep learning method is used to detect and generate the candidate region detection frames of the built-up areas on the large-scale remote sensing images,and then the SR model is used to accurately extract the built-up areas based on the candidate region detection frames to obtain the rural built-up areas on the large-scale remote sensing images.The relationship between different built-up areas sizes and convolution kernel in the SR model is analyzed and discussed through experimental tests.We compare this method with the anisotropic rotation-invariant textural measure(Pan Tex),Gabor algorithm and morphological building index(MBI)method,and discuss the influence of different sensors and different acquisition times on the extraction results.Finally we use the accuracy,recall and F-measure(F-value)to evaluate the extraction results.The following conclusions are drawn from the experiments:(1)The candidate rural built-up area is roughly localized by training enhanced Faster R-CNN model to get the candidate area detection frame,and the precise boundary of rural built-up area is obtained based on the candidate area detection frame by using SR model,and the accuracy evaluation of the three large-scale remote sensing image detection results finds that the accuracy can reach about 85%,the recall rate reaches more than 90%,and the F-value remains at 90%.The accuracy of extraction is high and compared with three different methods of extracting built-up area: Pan Tex,Gabor and MBI algorithm.The method is evaluated from both qualitative and quantitative aspects,and it is found that the method can accurately extract the boundary of built-up area,and the F-value can reach more than 91%,which is higher than the other three methods,so this method is feasible.(2)Two convolutional kernel sizes are considered in the SR model.The first kernel is a local average filter,which serves to perform mean filtering to obtain the average amplitude spectrum,and the second kernel is a Gaussian smoothing filter to smooth the final extraction results.The relationship between the size of the built-up area and the convolution kernel in the SR model is discussed,it is found that the size of the second kernel increases with the increase of the built-up areas' size,while the size of the first kernel does not change significantly with the increase of the built-up areas' size.By evaluating the proposed method,it is shown that the method can effectively and accurately perform fast extraction of different sizes of built-up area in remote sensing images.(3)The paper designed experiments to analyze and compare the influence of the acquisition time of images on the extraction results of built-up area.The results show that the difference between the lush vegetation and built-up area on the images in August is large,so the extraction effect of built-up area is better compared with that in April.Comparing the effects of different sensor images on the extraction results of built-up area,the experiment found that the SR method has no obvious preference for the selection of GF-1 and ZY-3 sensors in the experiment in the process of built-up area extraction,which can also verify the implementation of the method in this paper without the requirement of sensor types.