基于PRCUnet的高分遥感影像建筑物提取

High-Resolution Remote Sensing Image Building Extraction based on PRCUnet

基于高分辨率遥感影像的建筑物提取具有重要的理论与实际应用价值,深度学习因其优异的深层特征提取能力,已经成为高分影像提取建筑物的主流方法之一。本文在改进深度学习网络结构的基础上,结合最小外接矩形与Hausdorff距离概念,对建筑物提取方法进行改进。本文主要改进内容为:①基于Unet网络结构,利用金字塔池化模块(Pyramid Pooling Module,PPM)的多尺度场景解析特点,残差模块(Residual Block,RB)的特征提取能力以及卷积块注意力模块(Convolutional Block Attention Module,CBAM)对空间信息和通道信息的平衡能力。将金字塔池化、残差结构以及卷积块注意力模块引入到Unet模型中,建立PRCUnet模型。PRCUnet模型更关注语义信息和细节信息,弥补Unet对小目标检测的欠缺;②基于最小外接矩形与Hausdorff距离,改进建筑物轮廓优化算法,提高模型的泛化能力。实验表明,本文的建筑物提取方法在测试集上准确率、IoU、召回率均达到0.85以上,精度显著优于Unet模型,提取出的建筑物精度更高,对小尺度及不规则的建筑物有较好的提取效果,优化后的建筑物轮廓更接近真实的建筑物边界。

Building extraction based on high-resolution remote sensing images has important theoretical and practical applications.Deep learning has become one of the mainstream methods for extracting buildings from high-resolution images because of its excellent deep feature extraction ability.In this paper,based on an improved structure of deep learning network,we combined the concept of minimum outer rectangle and Hausdorff distance to improve the building extraction method.The main improvements in this paper are:①Based on the Unet network structure,we employed the multi-scale feature detection ability of Pyramid Pooling Module(PPM),the great feature extraction capability of Residual Block(RB),and the ability to balance spatial and channel information of Convolutional Block Attention Module(CBAM).The PPM,RB,and CBAM were introduced to the Unet model to build the PRCUnet model,which focuses more on semantic and detailed information and overcomes the limitation of Unet in small target detection;②We improved the building contour optimization algorithm based on the minimum outer rectangle and Hausdorff distance to improve the generalization ability of the model.Experiments show that the accuracy,IoU,and recall of the building extraction method proposed in this paper reached above 0.85 using the test set,significantly higher than those of the Unet model.The PRCUnet model also had better extraction effect on small-scale and irregular buildings than Unet,and the optimized building contours were close to the real building boundaries.