Spatialization of Building Data from High-Resolution Remote Sensing Images using CGSANet for Earthquake Emergency Response

With the development of remote sensing technology, the ability to acquire remote sensing images has been continuously improved and the application has become normalized, making remote sensing data an important source of building data [1]. In the field of earthquake emergency response, basic data for earthquake emergency response is an important foundation for rapid assessment of earthquake disasters [2-3], and building data is the most easily changed part [4]. Due to funding and technological constraints, it is difficult to collect and update building data. Currently, most of the data is obtained through surveys and statistics, which cannot reflect spatial distribution characteristics [5] and have poor timeliness. Spatialization of remote sensing images and building data is an effective method to solve these problems. The spatialization of building data is the extension of administrative unit statistical data in spatial scale, reflecting the spatial distribution status of buildings within administrative divisions and improving the spatial resolution of the data [6]. Extracting building vector data from high-resolution remote sensing images and spatializing them not only saves manpower and resources, alleviating the difficulties and workload of data collection, but also effectively improves data timeliness and provides better data support for rapid assessment of earthquake disasters [7]. \n\nMethods for automatic extraction of buildings from remote sensing images include traditional methods based on manually designed features and methods based on deep learning. Methods based on deep learning, especially Convolutional Neural Networks (CNN), have surpassed traditional methods in terms of speed and accuracy [8]. Deep neural network models have powerful feature representation and automatic learning capabilities, which can greatly overcome challenges such as large data volume, high dimensions, and complex scenes in remote sensing image data, and are widely used in automated building extraction [9]. Methods based on deep learning can be divided into region-based methods and end-to-end learning methods. Region-based methods represent low-level features such as shape, texture, and color of images using neural networks, use these features for semantic segmentation, and annotate the segmentation results with categories. Representative models include R-CNN (Region-based CNN) [10], including improved models such as Fast R-CNN [11], Faster R-CNN [12], and Mask R-CNN [13]. These methods rely on the selection of candidate regions. End-to-end training methods train networks using pixel-level ground truth of training samples to classify all pixels in an image. Most of these methods are based on the fully convolutional neural network (FCN) framework [14], including variants such as SegNet [15] and U-Net [16]. \n\nFCN is currently the mainstream network for building extraction, but it cannot capture global contextual information. U-Net fuses low-level positional information with semantic information by concatenating them, which can compensate for the spatial information loss caused by downsampling, but its ability to obtain multi-scale features is still imperfect [17]. The DeepLab series of models introduced the Atrous Spatial Pyramid Pooling (ASPP) module to fuse multi-scale contextual information, which can effectively capture multi-scale information. Based on this, Chen et al. [19] proposed a Contour-Guided and Local Structure-Aware Encoder-Decoder Network (CGSANet) model to effectively deal with building scale changes and obtain accurate building boundaries. This model preserves low-level spatial features related to building contours, obtains multi-scale contextual information through the ASPP module, and achieves complementary information between building edge semantic information and multi-scale regional semantic information. Its accuracy on the Wuhan University building dataset [20] can reach over 95%. \n\nMethods for spatializing building data mainly include data surveys based on individual buildings and simulation using large-scale housing statistics data. The former has a large workload and the latter has low data accuracy. Based on these two methods, scholars have proposed methods to establish housing distribution models using sampled housing data. Han Zhenhui et al. [21] combined population data spatialization with on-site survey data from kilometer grids to obtain the proportional relationship between housing types under different population densities and establish a housing data spatialization model. Ding Wenxiu et al. [22] conducted urban-rural and economic zoning in Sichuan Province and conducted grid sampling surveys based on zoning and land use data to establish a distribution model of housing structure types. Chen Jin et al. [23] used unmanned aerial vehicles, high-resolution satellite remote sensing images, and field survey data to obtain residential building type data by establishing interpretation criteria. \n\nTo improve the accuracy of building data and rapid assessment of earthquake disasters, while improving data timeliness, this paper takes Huangmei County, Hubei Province as the research area, uses high-resolution remote sensing images of Huangmei County and on-site sampled survey results as basic data, and uses the CGSANet model to extract building vector data from remote sensing images. By establishing a spatial distribution model of buildings at the regional scale through grid sampling, the spatialization of building data is achieved, and the accuracy of the spatialization results of building data is evaluated.