测绘通报 ›› 2022, Vol. 0 ›› Issue (6): 40-44.doi: 10.13474/j.cnki.11-2246.2022.0168.

• 学术研究 • 上一篇    下一篇

基于U-Net3+的高分遥感影像建筑物提取

窦世卿, 郑贺刚, 徐勇, 陈治宇, 苗林林, 宋莹莹   

  1. 桂林理工大学测绘地理信息学院, 广西 桂林 541006
  • 收稿日期:2021-07-07 发布日期:2022-06-30
  • 通讯作者: 徐勇。E-mail:yongxu@glut.edu.cn
  • 作者简介:窦世卿(1977-),女,博士,副教授,主要从事三维GIS与遥感技术应用的研究工作。E-mail:doushiqing@glut.edu.cn
  • 基金资助:
    广西八桂学者专项;国家自然科学基金(42061059);广西自然科学基金(2020GXNSFBA297160);桂林市科技局重点项目(20210128151428212);广西空间信息与测绘重点实验室资助(191851016)

Extraction of buildings with high-resolution remote sensing images based on U-Net3+ model

DOU Shiqing, ZHENG Hegang, XU Yong, CHEN Zhiyu, MIAO Linlin, SONG Yingying   

  1. College of Geomatics and Geo-information, Guilin University of Technology, Guilin 541006, China
  • Received:2021-07-07 Published:2022-06-30

摘要: 针对传统的高分影像建筑物提取方法存在分割精度低和分割边界模糊等问题,本文提出了一种基于U-Net3+模型的建筑地物语义分割方法。该模型以U-Net网络结构为基础,首先使用全尺度的跳跃连接将不同尺度的特征图相融合;然后通过深度监督从多尺度聚合的特征图中学习特征表达,并使用交叉熵损失函数进行训练;最后根据数据集特征,调试出不同的模型参数并以此模型进行测试,以达最佳的分割效果。试验结果表明,与U-Net和U-Net++模型相比,基于该方法的影像分割精度及地物边缘分割完整度均得到了显著提升,且当设置历元为15时,精度最高。使用该方法对高分辨率遥感影像中建筑物进行的分割试验,精度达96.62%,平均交并比(mIoU)达0.902 7,并减少了错分、漏分,同时也减少了模型参数,模型损失收敛速率快且缩短了训练周期,显著提升了建筑物提取精度。

关键词: 地物信息提取, U-Net3+模型, 全尺度跳跃连接, 深度监督, 精度

Abstract: In response to the problems of low segmentation accuracy and blurred segmentation boundaries in traditional methods for extracting buildings from high-resolution remote sensing images, this paper proposes a semantic segmentation method for building features based on the U-Net3+ model. Firstly, on the basis of the U-Net network structure, the feature maps of different scales are fused using full-scale jump connection.Then, the feature expressions are learned from the multi-scale aggregated feature maps by deep supervision, and the cross-entropy loss function is used for training.Finally, different model parameters are tuned and tested according to the dataset characteristics to achieve the best segmentation effect. Experimental results show that the image segmentation accuracy and feature edge segmentation completeness based on the U-Net3+ model significantly improve in comparison with the U-Net and U-Net++ models, and the highest accuracy is achieved when setting epoch as 15 in all three models. Based on the U-Net3+ model, the segmentation accuracy of building features for high-resolution remotesensing images reaches 96.62% and the average intersection ratio of mIoU reaches 0.902 7, which reduces the phenomenon of missegmentation and omission, and reduces the model parameters, the model loss convergence rate is fast and the training period is shortened, and the extraction accuracy of buildings is significantly improved.

Key words: feature information extraction, U-Net3+ model, full-scale jump connectivity, deep supervision, accuracy

中图分类号: