基于改进Encoder-Decoder网络的遥感影像道路提取

A Remote Sensing Image Road Extraction Method Based on Improved Encoder-Decoder Network

摘要: 本文针对高分遥感影像中蕴含的丰富信息给道路提取结果带来的干扰问题，提出了一种基于改进Encoder-Decoder网络的高分遥感影像道路提取方法。首先在编码区引入残差模块提取图像特征信息，然后在网络的中心区域引入空洞卷积模块，进一步拓展了识别道路特征像素信息的感受野，并保障了特征图分辨率以及像素点空间信息保持不变，增强了网络的细节提取能力，最后使用Sigmoid函数对特征图进行分类。本次实验采用马萨诸塞州道路数据集作为训练数据，实验结果表明，本文所提出的基于改进编码–解码网络道路提取方法将整体的精度、召回率以及F1-score指标分别提升至91%，58%和71%，与U-Net模型相较有着明显的提升。

Abstract: Aiming at the rich details and simple semantics of road information in high-resolution remote sensing images, a high-resolution remote sensing image road extraction model based on improved Encoder-Decoder network is designed. The residual module is introduced in the encoding area to extract image feature information. The introduction of dilated convolution module in the central area of the network further expands the receptive field of identifying the characteristic pixel information of the road, and ensures that the resolution of the characteristic map and the spatial information of the pixel remain unchanged, which enhances the detail extracting ability of the network. Finally, the Sigmoid function is used to classify the feature map. Through the comparison test of the experimental verification set, the overall accuracy, recall rate and F1-score index of the road extraction method based on the improved encoding-decoding network proposed in this paper reached 91%, 58%, and 71% respectively. Comparing to the U-Net model, there is a significant improvement.