基于多尺度特征融合的U-Net网络高分影像不透水面提取研究

不透水面作为监测城市生态环境的重要指标，其信息提取具有重要意义。由于城市地表的复杂性及细化的城市管理需要，急需提取高精度的城市不透水面。但是基于传统方法提取高精度的城市不透水面面临巨大困难。而深度学习方法因其自动化提取影像特征的特点逐渐成为遥感影像地物提取的新兴方法。基于此，采用多尺度特征融合的U-Net深度学习方法以提升语义分割精度，开展高分辨率遥感影像不透水面的精确提取研究。模型引入残差模块代替普通卷积以加深网络，提取更多影像特征；加入金字塔池化模块增强网络对复杂场景的解析能力；利用跳跃连接方式融合不同尺度特征，有利于恢复空间信息。以广州市航摄正射影像为数据源，通过卷积神经网络将遥感影像分割为背景、其他、植被、道路和房屋5种地物类型，将其与人工目视解译的地面真值进行验证，最终提取研究区域不透水面。实验证明：多尺度特征融合的U-Net模型总体精度和Kappa系数分别为87.596% 和0.82。在定性与定量两个方面均优于传统的监督分类法、面向对象分类法和经典U-Net模型法。结果表明：该模型利用多维度影像特征信息，有效提升了复杂场景图像的分割精度，分割效果好，适用于高分辨率遥感影像不透水面提取，该研究成果可为城市环境监测提供数据支撑。

Impervious Surface Extraction from High-resolution Images based on Multi-scale Feature Fusion in U-Net Network

As an important index for monitoring urban ecological environment， the extraction of impervious surface is of great significance. Due to the complexity of urban surface and the need of detailed urban management， it is urgent to extract high-precision urban impervious surface. However， it is difficult to extract high precision impervious surface based on traditional methods. Deep learning method has gradually become a new method of remote sensing image feature extraction because of its characteristics of automatic image feature extraction. Based on this， this paper uses the U-Net deep learning method based on multi-scale feature fusion to improve the semantic segmentation accuracy， and carries out the research on the accurate extraction of impervious surface from high resolution remote sensing images.The residual module is introduced instead of convolution to deepen the network and extract more image features， the pyramid pooling module is added to enhance the network's ability to resolve complex scenarios. It is beneficial to recover spatial information by combining different scale features with jump connection. In this paper， aerial orthophoto images of Guangzhou were taken as the data source. Through convolutional neural network， the remote sensing image is segmtioned into five types of features： background， others， vegetation， road and building. Verify it with the ground truth value of manual visual interpretation， finally， the impervious surface of the study area was extracted. Experiments show that the overall accuracy and Kappa coefficient of the U-NET model are 87.596% and 0.82， respectively. It is superior to traditional supervised taxonomy， object-oriented taxonomy and classical U-Net model in both qualitative and quantitative aspects. The results show that the model can effectively improve the segmentation accuracy of complex scene images by using the multi-dimensional image feature information， and the segmentation effect is good， which is suitable for the extraction of impervious water from high resolution remote sensing images. The research results in this paper can provide data support for urban environmental monitoring.