基于深度学习的高分二号影像农村道路提取

Rural Road Extraction from GF-2 Image Based on Deep Learning

针对高分遥感影像在农村道路提取中的研究和应用少,提取结果存在边界不清晰和道路不完整等问题,本文以高分二号影像为数据源,提出一种改进的DeepLabv3+深度学习模型,优选MobileNetV2作为特征提取网络,并在MobileNetV2反向残差块中插入坐标注意力模块,使得网络能捕捉到具有精确位置的信息,同时在空间金字塔池化后加上通道注意力机制,使用多尺度注意力机制对输出特征进行合并,将注意力集中在信息量更大的特征上,避免使用多个相似的特征图。实验结果表明:改进后的DeepLabv3+准确率达85.74%,召回率为83.21%,F 1评分为0.84,相比原始的DeepLabv3+模型,各精度指标都有一定的提升。本研究可为农村道路的高精度提取提供一定的技术支持。

In view of the lack of research and application of high-resolution remote sensing images in rural road extraction,the extraction results have some problems such as unclear boundaries and incomplete roads.In this paper,taking GF-2 image as the data source,an improved DeepLabv3+deep learning model is proposed.MobileNetV2 is selected as the feature extraction network,and the coordinate attention module is inserted into the reverse residual block of MobileNetV2,so that the network can capture the information with accurate position.At the same time,channel attention mechanism is added after the spatial pyramid pool,and multi-scale attention mechanism is used to merge the output features,focusing on the features with more information and avoiding using multiple similar feature maps.The experimental results show that the accuracy of the improved DeepLabv3+model is 85.74%,the recall rate is 83.21%,and the F1 score is 0.84.Compared with the original DeepLabv3+model,each precision has been improved to some extent.This study can provide some technical support for high-precision extraction of rural roads.