Research On Building Extraction Method Of High Resolution Remote Sensing Images Based On Morphology And Deep Learning

Metropolises are a vital assurance for people’s livelihoods and industry.The emergence and implementation of the smart city concept in recent years have heightened the demand for sophisticated urban system services,in order to attain refined management and dynamic information updates.During the ongoing evolution of China’s satellite remote sensing sector,remote sensing technology has demonstrated significant potential in the intelligent extraction of urban structures.The comprehensive and intricate data within high-spatial resolution optical remote sensing imagery is particularly conducive to precise building extraction.Accordingly,GF-2and Jilin-1 high-resolution satellite remote sensing images were selected as data sources to investigate the suitable extraction techniques for urban structures,and the specific research content and findings are as follows:(1)A method for extracting buildings based on an adaptive morphological index is proposed.The approach involves improvements based on the analysis of standard morphological building indices.Initially,the RGB color space is transformed into the YCb Cr color space,leveraging the color space transformation theory to enable comprehensive mining of color information.Subsequently,adaptive multi-scale top hat transformation is utilized to determine the optimal scale space associated with the building and extract relevant feature information from the image.Finally,the extracted results are subjected to post-processing through vegetation index constraints and area suppression.The experiment was conducted using remote sensing image data obtained from the GF-2 satellite and the Jilin-1 satellite.The experimental results demonstrate that the precision of this method can reach 72.5% and 77.05%,respectively,and the overall classification accuracy reaches 88.6% and 88.43%,respectively.(2)A building extraction method based on the Refine-UNet network model is presented.The proposed approach leverages the characteristics of the U-shaped network architecture and introduces the Refine-UNet network model to achieve highprecision building extraction.The model consists of three components: the encoder module,the decoder module,and the refined hopping connection module,with the latter comprising an Atrous Spatial Pyramid Pooling(ASPP)module and four improved Depthwise Separable Concrete(IDSC)modules.Using remote sensing image data obtained from the Jilin-1 satellite and comparing it with four deep learning networks,namely UNet,Seg Net,Deep Lab V3+,and PSPNet,the experimental results demonstrate that the Refine-UNet network achieves a building extraction accuracy of95.1% and an Intersection over Union(Io U)score of 87%,which is 6.4% and 8.5%higher than that of the UNet network,respectively.(3)This thesis presents building extraction methods that combine joint constraint-enhanced morphological index and deep learning.Leveraging the principle of the attention mechanism,the proposed approach utilizes the building extraction results obtained from standard morphological index and adaptive morphological index as constraint windows to linearly enhance the original remote sensing image data.Using remote sensing images of buildings obtained from the Jilin-1 satellite,the impact of combining morphology and deep learning on the accuracy of building extraction was extensively analyzed.Utilizing high-resolution satellite remote sensing imagery as the primary data source,this study investigates the impact of conventional morphological indices,deep learning algorithms,and their combination on the precise extraction of building targets.The proposed method is experimentally validated,affirming its rationality for efficiently leveraging building information logic in images.This research serves as an effective foundation for future urban and rural development,as well as smart city planning and construction.