基于特征一致性感知的遥感影像道路提取方法

Road Extraction Method from Remote Sensing Images with Feature Consistency Perception

道路提取是遥感信息提取中的一个重要课题。然而,当道路被建筑物和树木遮挡时,现有的道路提取方法在感知道路特征的全局一致性方面较弱,导致道路提取结果破碎化。为了解决这个问题,提出特征增强和特征一致性感知网络(FECP-Net),该网络由初始道路提取网络(CRE-Net)以及特征增强和特征一致性感知(FECP)模块组成。在该网络中,CRE-Net用于提取初始道路信息和特征,FECP模块通过将粗略的道路信息与不同尺度的道路特征相连接来增强道路特征的一致性,提高道路提取结果的完整性。在CHT数据集、马萨诸塞州数据集和DeepGlobal数据集上,将所提方法与DGRN、U-Net和D-LinkNet等不同方法进行对比。在马萨诸塞州数据集上的结果表明,与其他方法相比,所提方法的交并比(IOU)提高了0.45百分点、3.36百分点、9.48百分点,F1分数提高了1.26百分点、2.76百分点、8.12百分点,召回率提高了4.60百分点、5.93百分点、12.46百分点。所提方法可以提取更完整的道路,改善道路破碎化和提取结果不连通的现象。

Road extraction is an important topic in remote-sensing information extraction.However,for cases when buildings and trees obstruct roads,existing road extraction methods have a weak global consistency in sensing road features,resulting in fragmented road extraction results.A feature enhancement and consistency awareness network(FECP-Net)is proposed to address this issue.The network comprises an initial road extraction network(CRE-Net)and a feature enhancement and consistency awareness(FECP)module.In this network,CRE-Net extracts the initial road information and features.In contrast,the FECP module enhances the consistency of road features.It improves the completeness of road extraction results by connecting rough road information with road features of different scales.The proposed method was compared with other methods,namely,DGRN,U-Net,and D-LinkNet,on the CHT,Massachusetts,and DeepGlobal datasets.The results on the Massachusetts dataset showed that compared to other methods,the proposed method increased the intersection over union(IOU)by 0.45 percentage points,3.36 percentage points,and 9.48 percentage points,respectively,the F1 scores increased by 1.26 percentage points,2.76 percentage points,and 8.12 percentage points,respectively,and the recall rates increased by 4.60 percentage points,5.93 percentage points,and 12.46 percentage points,respectively.The proposed method can extract the information of more complete roads and improve road fragmentation and disconnection extraction results.