基于深度学习模型的城市高分辨率遥感影像不透水面提取

不透水面是衡量城市生态环境状况的重要指标。城市土地利用的复杂性和不透水表面材料的多样性,导致直接从高分辨率遥感影像中提取不透水表面具有挑战性。针对城市尺度高分辨率遥感影像的不透水面提取要求,本文提出基于深度学习的城市不透水面提取模型。首先,利用深度卷积神经网络对影像特征进行提取;然后,根据其邻域关系构建概率图学习模型,进一步引入高阶语义信息对特征进行优化,实现不透水面的精确提取。本文选取武汉市为实验区,以高分二号卫星遥感影像作为数据源,完成了不透水面专题信息提取,其中自动提取准确率在建成区为89.02%、在城乡结合部为95.55%。与随机森林（RF）和支持向量机（SVM）等经典方法对比,结果表明深度学习不透水面提取方法有较高的提取精度和细节准确性,建成区的总体精度相比于RF和SVM算法分别提升2.18%和1.68%。最后,对武汉市各主要行政区不透水面信息进行统计和分析,结果表明其中江汉区和武昌区2个核心主城区不透水面占比超过60%,并对武汉市现状和发展规划特点进行了讨论。本文研究成果可为海绵城市和生态城市的建设提供基础技术支撑和数据参考。

Extraction of Urban Impervious Surface from High-Resolution Remote Sensing Imagery based on Deep Learning

Impervious surface is an important indicator of urban ecological environment, which is of great significance for urbanization and environmental quality assessment. The complexity of urban land use and the diversity of impervious surface materials make it a challenge to extract impervious surface directly from high-resolution remote sensing imagery. To meet the requirement of impervious surface extraction from high-resolution remote sensing imagery at the urban scale, a model of impervious surface extraction based on deep learning was proposed in this paper. Firstly, deep convolution neural network was used to extract image features. In extracting impervious surface of complex cities, a convolution layer and a pool layer were retained. While the void convolution was introduced to increase the field of receptivity and reduce the loss of information, so that each convolution output contained a larger range of information. Secondly, a probabilistic graph learning model was constructed according to its neighborhood relationship, and high-order semantic information was introduced to optimize the features to achieve accurate extraction of impervious surfaces. This paper choosed Wuhan as the experimental area, and took GaoFen-2 satellite remote sensing imagery as the data source to implement the proposed model for the extraction of impervious surface thematic information. The automatic extraction accuracy was 89.02% in the construction area and 95.55% in the urban-rural junction. Compared with the traditional machine learning algorithms such as random forest and support vector machine, the efficiency and accuracy of the proposed deep learning method were better. Statistics and analysis of the impervious surface information of the main administrative regions in Wuhan showed that the proportion of impervious surface in the whole territory of Wuhan was 11.43%, and the proportion of impervious surface in the core main urban area was close to 70%. Additionally, the present situation and development planning characteristics of Wuhan were analyzed and discussed. The impervious surface can be used as a link between urban development level and environmental quality. The distribution of impervious surface in Wuhan development planning of various administrative districts is closely related to the sustainable development of the city. Our findings suggest that the deep learning method is effective for the extraction of impervious surfaces from high-resolution remote sensing imagery, and can provide technical support and data reference for the construction of sponge city and ecological city.