基于改进U-Net模型的遥感影像道路提取方法研究

Research on Road Extraction Method from Remote Sensing Image Based on Improved U-Net Model

针对高分辨率遥感影像背景信息复杂,道路提取难度大,自动化程度低等问题,论文提出了一种改进的U-Net的道路提取方法。首先,编码器使用VGG16网络结构替代原始U-Net编码器结构;然后,在每个编码器和解码器块后加入特征压缩激活模块(SENet)增强网络特征学习能力;最后,使用Dice损失函数和二分类交叉熵损失函数复合的损失函数进行训练,减轻了道路提取任务中的样本不平衡问题。在Massachusetts Road数据集上的结果表明,改进后的算法对道路提取结果得到了有效的提升。所提方法在测试集上的精确度、召回率、F1-score和mIoU评价指标分别达到82.5%、77.8%、80.0%及82.1%,在测试影像中对错综交叉的道路具有更好的识别效果。

In view of high-resolution remote sensing images being with complex background information,road extraction in the high-resolution remote sensing images being difficult and having a low degree of automation,this paper proposes an improved U-Net road extraction method.First,the VGG16 network structure is employed in the encoder to replace the original U-Net encoder structure,then,a feature compression activation module(SENet)is added after each encoder and decoder block to enhances the ability of network feature learning.Finally,the loss function combined with the Dice loss function and the binary cross-entropy loss function is used for training,which reduces the sample imbalance problem in the road extraction task.The experimental results on the Massachusetts Road data set show that the improved algorithm has effectively improved the road extraction results.The preci-sion,recall,F1-score and mIoU evaluation indicators of the proposed method on the test set reached 82.5%,77.8%,80.0%,and 82.1%,respectively.In the test image,it has a better recognition effect on roads with different widths and irregular shapes.