基于注意力增强全卷积神经网络的高分卫星影像建筑物提取

Building extraction using high-resolution satellite imagery based on an attention enhanced full convolution neural network

从卫星遥感影像中自动提取建筑物在国民经济社会发展中具有广泛的应用价值,由于卫星遥感影像存在地物遮挡、光照、背景环境等因素的影响,传统方法难以实现高精度建筑物提取。采用一种基于注意力增强的特征金字塔神经网络方法(FPN-SENet),利用多源高分辨率卫星影像和矢量成果数据快速构建大规模的像素级建筑物数据集(SCRS数据集),实现多源卫星影像的建筑物自动提取,并与常用的全卷积神经网络进行对比。研究结果表明:SCRS数据集的提取精度接近国际领先的卫星影像开源数据集,且假彩色数据精度高于真彩色数据;FPN-SENet的建筑物提取精度优于其他常用的全卷积神经网络;采用交叉熵和dice系数之和为损失函数能够提升建筑物提取精度,最好的分类模型在测试数据上的分类总体精度为95.2%,Kappa系数为79.0%,F1分值和IoU分别达到了81.7%和69.1%。该研究可为高分辨率卫星影像建筑物自动提取提供参考。

Automatic extraction of buildings from satellite remote sensing images has a wide range of applications in the development of economy and society.Due to the influence of mutual occlusion,illumination,background environment and other factors in satellite remote sensing images,it is difficult for traditional methods to achieve high-precision building extraction.This paper proposes an attention enhanced feature pyramid network(FPN-SENet)and constructs a large-scale pixel-wise building dataset(SCRS dataset)by using multi-source high-resolution satellite images and vector data to realize the automatic extraction of buildings from multi-source satellite images,and compares it with the other full convolution neural networks.The results show that the accuracy of building extracted from SCRS dataset is close to the world’s leading open source satellite image dataset,and the accuracy of Pseudo color data is higher than that of true color data The accuracy of FPN-SENet is better than that of other full convolution neural networks.The extraction of building can also be improved by using the sum of cross entropy and Dice coefficient as the loss function.The overall accuracy of the best classification model is 95.2%,Kappa coefficient is 79.0%,and F1-score and IoU are 81.7%and 69.1%respectively.This study can provide a reference for building automatic extraction from high-resolution satellite images.