基于深度学习的遥感影像道路提取

Deep learning based road extraction from remote sensing images

针对遥感影像道路提取任务中因地物环境复杂而导致道路掩码提取精度差、道路中心线断裂和不连续的现象,提出一种深度学习语义分割模型CP-Unet进行道路掩膜提取,将提取的道路掩膜进行形态学处理,采用ZS细化算法进行道路中心线提取,并提出一种优化连接模型进行中心线处理。CP-Unet扩大了模型感受野,增强了对道路信息的捕捉能力和融合能力,提高了道路掩码的提取精度,优化连接模型通过设定几何约束条件进行中心线断点连接,提高道路中心线的连续性。以新疆某团场分辨率为0.5 m的wordview3卫星影像为实验数据,实验结果表明:CP-Unet的精确率、召回率、平均交并比分别提高到91.92%、88.27%和81.43%,能够较好地克服复杂环境干扰,提取精度较高,中心线提取方法在两种不同的复杂环境下提取准确率和完整率分别为94.82%、92.79%和96.77%、94.17%,提取结果更加连续且完整。

Due to the complex ground object environment,the road mask extraction accuracy is poor,and the road center line is fractured and discontinuous in the road extraction task by remote sensing images.Therefore,a semantic segmentation model CP⁃Unet based on deep learning is proposed for road mask extraction.The extracted road mask is processed by morphology,and the ZS refinement algorithm is adopted for road center line extraction.An optimized connection model is proposed for center line processing.CP⁃Unet expands the model field of sensitivity,enhances the capturing ability and fusion ability of road information,improves the extraction accuracy of road mask,and optimizes the connection model.The center line breakpoint is connected by setting geometric constraints,so as to improve the continuity of road center line.The wordview3 satellite image with a field resolution of 0.5 m in Xinjiang is taken as the experimental data.The experimental results show that the accuracy rate and recall rate of CP⁃Unet are improved to 91.92%and 88.27%,respectively,and its average intersection ratio is increased to 81.43%,the CP⁃Unet can overcome the interference of complex environment and reach high extraction accuracy rate,the accuracy rate and integrity rate of the center line extraction method in two different complex environments are 94.82%and 92.79%,96.77%and 94.17%,respectively,and its extraction results are more continuous and complete.