改进全卷积网络方法的高分二号影像农村道路提取

针对目前利用高分遥感数据提取农村道路的研究与应用少，提取结果精准度不够的问题，提出了结合空洞卷积和ASPP（Atrous Spatial Pyramid Pooling）结构的改进全卷积农村道路提取网络模型DC-Net（Dilated Convolution Network）。该模型基于全卷积的编解码结构来提取道路深度特征信息，同时针对农村道路细长的特点，在解编码层之间加入了以空洞卷积为基础的ASPP（Atrous Spatial Pyramid Pooling）结构来提取道路的多尺度特征信息，在不牺牲特征空间分辨率的同时扩大了特征感受野FOV（Field-of-View），从而提高细窄农村道路的识别率。以长株潭城市群郊区部分区域为试验对象，以高分二号国产卫星遥感影像为实验数据，将本文提出的方法与经典的几种全卷积网络方法进行实验结果对比分析。实验结果表明：（1）本文所提出的道路提取模型DC-Net在农村道路的提取上具有可行性，整体提取平均精度达到98.72%，具有较高的提取精度；（2）对比几种经典的全卷积网络模型在农村道路提取上的效果，DC-Net在农村道路提取的精度和连结性、以及树木和阴影的遮挡方面，均表现出了较好的提取结果；（3）本文提出的改进全卷积网络道路提取模型能够有效地提取高分辨率遥感影像中农村道路的特征信息，总体提取效果较好，为提高基于国产高分影像的农村道路提取精度提供了一种新的思路和方法。

Road extraction in rural areas from high resolution remote sensing image using a improved Full Convolution Network

Aiming at the problems of limited research, application of extracting rural roads with high-resolution remote sensing data, and insufficient accuracy of extraction results, a new improved full convolution rural road extraction network model Distributed Convolution network (DC net) is proposed; it combines void convolution and Air Spatial Pyramid Pooling (ASPP) structure. The model extracts the depth feature information of the road based on the full convolution encoding and decoding structures. At the same time, in accordance with the characteristics of the slender rural roads, the ASPP structure based on the hollow convolution is added between the decoded layers to extract the multiscale characteristic information of the road, and the Field of View (FOV) is expanded without sacrificing the spatial resolution of the feature, thereby improving the recognition rate of narrow and fine rural roads. Some suburban areas of Changzhutan city group are considered the experimental objects and the domestic satellite remote sensing image of Gaogaoer as the experimental data. Experimental results are compared with those of the classical methods of all convolution networks. The results show that: (1) the proposed road extraction model DC net is feasible in rural road extraction, with the overall extraction average accuracy reaching 98.72%, indicating high extraction accuracy; (2) comparative results of the effect of several classic full convolution network models on rural road extraction, DC net extraction accuracy and connectivity, as well as tree and shadow shading in the aspect of block are acceptable; (3) the improved road extraction model of the entire proposed convolution network can effectively extract the feature information of rural roads in high-resolution remote sensing images. The overall extraction effect is improved; it provides a new approach for improving the precision of rural road extraction based on domestic high-resolution images.Based on the full convolution network model in deep learning, this paper proposes an improved full convolution rural road extraction model DC net which combines hole convolution and ASPP structure. According to the characteristics of long and thin and connectivity of rural roads, this method combined with hole convolution to expand the receptive field of feature map in the process of model training, which makes the extraction of rural roads more complete.