Research On Semantic Segmentation Method Of Buildings In Remote Sensing Images Based On Deep Learning

Buildings are an important part of human life,and high precision extraction of buildings in remote sensing images can provide effective help for military operations,population distribution surveys,and land planning.At present,convolutional neural networks have been applied in the field of semantic segmentation of high-resolution visible remote sensing images,but when there are objects with the same scale as small buildings,mis-segmentation and omission may occur;when the scale of buildings is too small,problems such as incorrect extraction may occur.To address the above problems,after analyzing and studying the related literature on building extraction in domestic and foreign remote sensing images,the main work of the paper is as follows.Firstly,an improved U-Net network structure is proposed to improve the extraction accuracy by adding a compression activation module to filter effective features and a ASPP module to enrich the building features extracted by the network.The improved U-Net network is compared with other models on Massachusetts and WHU building datasets,and the results show that the improved U-Net network has improved the accuracy,Io U and F1-score compared with the U-Net,Seg Net and PSPnet models.Second,an encoder-decoder network is proposed based on the improved U-Net network structure.An improved Dense-ASPP module is added to the encoder to enhance building feature extraction at different scales;a convolutional block attention module is added to strengthen building features and enhance the network’s ability to extract buildings in complex backgrounds.A multi-scale feature fusion layer is added to the decoder to solve the extraction difficulties caused by small inter-class differences.The comparison experimental results show that this encoder-decoder network outperforms the U-Net,Seg Net and PSPnet models in terms of accuracy,Io U,and F1-score on both Massachusetts and WHU building datasets,and further improves over the improved U-Net network in this paper.Finally,the paper further verifies the effects of adding different modules and improvements to the encoder-decoder on improving the building map extraction by ablation experiments.The experimental results all show that the method proposed in this paper solves the problems of misclassification,omission and incorrect extraction of small buildings to a certain extent,and improves the extraction accuracy of buildings.