基于改进U-Net的遥感影像建筑物提取方法

Building Extraction from Remote Sensing Images Based on Improved U-Net

提取遥感影像中的建筑物对智慧城市建设有着重要意义。针对传统方法提取背景复杂遥感影像时出现的精度低等问题,提出一种基于U-Net的遥感影像建筑物提取方法(MA-Unet)。该方法主要由编码器和解码器组成。在编码器中,引入卷积块注意力模块,其中通道注意力模块用来筛选更重要的特征,抑制无效特征,空间注意力模块用来筛选更深层次的语义特征,引入空洞空间金字塔池化模块提取不同尺度的特征。在解码器中,为了融合不同尺度大小的物体特征,将解码器中的特征图上采样后进行串联,这种信息聚合在某种程度上解决了不同尺度物体检测困难的问题。实验结果表明:MA-Unet方法在Massachusetts建筑物数据集上的准确率、精度、交并比(IoU)分别优于U-Net网络1.7个百分点、2.1个百分点、1.6个百分点,在WHU建筑物数据集上的准确率、精度、IoU分别优于U-Net网络1.1个百分点、1.4个百分点、2.3个百分点,是一种更为有效且具有一定实际应用价值的目标提取方法。

Building extraction from remote sensing images is of great significance to the construction of smart cities.Aiming to improve the low accuracy of traditional methods in extracting remote sensing images with a complex background,a remote sensing image building extraction method(MAUnet)based on UNet is proposed.This method mainly uses an encoder and a decoder.A convolutional block attention module is introduced into the encoder,in which a channel attention module is used to screen more important features and suppress invalid features,and a spatial attention module is used to screen deeper semantic features.An atrous spatial pyramid pooling module is introduced to extract features with different scales.In the decoder,to fuse object features with different scales,feature maps in the decoder are upsampled and connected in series.This information aggregation solves the difficulty of detecting objects with different scales to some extent.The experimental results show that MAUnet method is superior to the UNet method in terms of accuracy,precision,and intersection over union(IoU)by 1.7 percentage points,2.1 percentage points,and 1.6 percentage points on the Massachusetts building dataset and by 1.1 percentage points,1.4 percentage points,and 2.3 percentage points on the WHU building dataset,respectively.It is a more effective and practical target extraction method.