基于FPN Res-Unet的高分辨率遥感影像建筑物变化检测

针对城市土地资源变化检测工作繁杂、工作量大、自动化程度低等问题,本文提出一种基于深度学习模型的高分辨率遥感影像建筑物变化检测方法,将语义分割的思想引入到遥感变化检测。基于残差结构特征较卷积层提取性能更优和特征金字塔网络多尺度预测的特点,将残差结构和特征金字塔网络融合到Unet模型中,建立FPN Res-Unet模型。该模型以Unet为基础,引入ResNet18的残差结构作为编码路径特征提取层,在每次卷积后使用边界填充,使得输入图像和输出图像尺寸一致;在解码路径每级上采样过程中,拓展支路径将特征金字塔网络融合到模型的网络主干中,将残差结构、Unet及特征金字塔网络的优点相互融合,增强了Unet的特征提取,弥补了语义分割网络对小目标检测的欠缺;在获取深层语义信息的同时关注细节信息,提高建筑物变化检测精度。实验表明,该方法在所用数据集,准确率、召回率、F1 3种指标均达到90%以上。

Building Change Detectionin High Resolution Remote Sensing Images Based on FPN Res-Unet

In view of the complexity, heavy workload,and low degree of automation in current survey of land resource change detection, a building change detection method of high-resolution remote sensing image based on deep learning model is proposed, and the idea of semantic segmentation is applied to change detection. Based on the better extraction performance of the residual structure than convolution layers and the characteristics of multi-scale prediction of feature pyramid networks, the residual structure and FPN are fused into Unet model to establish FPN Res-Unet. The model is based on Unet with ResNet residual structure as its feature extraction layer. After each convolution, padding is used to keep the size of the input image and the output image consistent. In the process of sampling at each level of the decoding path, the branch path is expanded to fuse FPN into the network trunk of the model. It fully combines the advantages of residual structure, Unet and FPN, which makes it pay attention to details while obtaining deep semantic information, and improves the detection accuracy of building change. Experiments show that the accuracy rate, recall rate and F₁ of the method in the data set used reach more than 90%.