基于八方向条状池化的遥感影像道路提取方法

Road Extraction from Remote Sensing Images Based on Eight-directional Stripe Pooling

针对在卫星遥感图像中道路提取存在云雾植被遮挡、分辨率低等客观条件导致提取精度低,利用道路具有带状、连通性等特征,提出了一种基于八方向条状池化的遥感影像道路提取方法(DLinkNet-Road)。首先,结合道路的多方向带状特征和条状池化提取细长目标的优势,构建了八方向条状池化道路提取模块,有效建立了道路像素长距离多方向依赖关系。其次,考虑到遮挡等导致道路断裂以及卷积池化操作导致道路轮廓细节信息丢失的问题,设计了道路特征加权补偿模块,并构建了加权特征融合结构,有效融合了多个尺度特征的道路信息。在DeepGlobe和Massachusetts两个道路数据集实验,本文方法的交并比(intersection over union,IoU)分别达到67.42%和66.38%,相较于基线模型提高了3.89%和3.17%。实验结果表明,所提模型能改善道路提取中的断线现象,保证道路提取结果的完整性。

In order to address the challenges posed by cloud and vegetation occlusion,as well as low resolution in satellite remote sensing images,which often result in low extraction accuracy,a road extraction method of DLinkNet-Road based on eight-directional stripe pooling was proposed for road extraction in satellite remote sensing images.Firstly,by combining the multi-directional strip features of roads and the advantages of strip pooling for extracting slender targets,an eight directional strip pooling road extraction module was constructed,effectively establishing long-distance multi-directional dependencies of road pixels.Secondly,considering the problems of road breakage caused by occlusion and loss of road contour details caused by convolution pooling operations,a road feature weighted compensation module was designed and a weighted feature fusion structure was constructed to effectively fuse road information from multiple scale features.In experiments on two road datasets,DeepGlobe and Massachusetts,the proposed method achieves an intersection over union(IoU) ratio of 67.42% and 66.38%,respectively,which was 3.89% and 3.17% higher than the baseline model.The experimental results show that the model proposed can improve the disconnection phenomenon in road extraction and ensure the integrity of the road extraction results.