10.3969/j.issn.1003-5060.2025.01.005

基于 EfficientNet 深度学习网络的遥感影像地物分割方法

姜克儒 $ ^{1} $，刘军 $ ^{2} $，谢枫 $ ^{3} $，盛金马 $ ^{1} $，刘耀中 $ ^{3} $，许水清 $ ^{4} $

（1. 国网安徽省电力有限公司经济技术研究院，安徽合肥 230022；2. 国网安徽省电力有限公司建设分公司，安徽合肥 230071；3. 中国能源建设集团安徽省电力设计院有限公司，安徽合肥 230093；4. 合肥工业大学电气与自动化工程学院，安徽合肥 230009）

摘要

在电网工程建设中, 自动提取遥感影像中包含的地物信息对实现电力规划自动化具有重要意义。文章提出一种基于深度学习的遥感影像地物分割方法, 为解决遥感影像目标丰富、尺度多样的问题, 以 UNet 网络为基本架构, 选择 EfficientNet 网络作为主干网络, 并加入特征融合; 为解决遥感影像类别不均衡和泛化能力问题, 采用联合损失函数和先进的数据增强方法。实验结果表明, 所提方法能够有效提高遥感图像地物分割精度, 对小目标和大目标均有较好的分割效果。

关键词

遥感图像；地物分割；深度学习；特征融合；损失函数

中图分类号：TP751.1

文献标志码：A

文章编号：1003-5060（2025）01-0032-05

Ground object segmentation method for remote sensing images based on EfficientNet deep learning network

JIANG Keru $ ^{1} $, LIU Jun $ ^{2} $, XIE Feng $ ^{3} $, SHENG Jinma $ ^{1} $, LIU Yaozhong $ ^{3} $, XU Shuiqing $ ^{4} $

(1. Economic and Technical Research Institute, State Grid Anhui Electric Power Co., Ltd., Hefei 230022, China; 2. Construction Company, State Grid Anhui Electric Power Co., Ltd., Hefei 230071, China; 3. China Energy Engineering Group Anhui Electric Power Design Institute Co., Ltd., Hefei 230093, China; 4. School of Electrical Engineering and Automation, Hefei University of Technology, Hefei 230009, China)

Abstract

In power grid project construction, automatic extraction of ground object information from remote sensing images is very important for realizing automation of power planning. In this paper, a ground object segmentation method for remote sensing images based on deep learning is proposed. In view of the multiple targets and diversified scales of remote sensing images, the UNet is used as the basic framework, EfficientNet is selected as the backbone network, and feature fusion is added. In order to solve the problem of classification unbalance and generalization ability of remote sensing images, combined loss function and advanced data enhancement method are adopted. Experiments show that the proposed method can effectively improve the ground object segmentation accuracy of remote sensing images, and has good segmentation effect for both small and large targets.

Keywords

remote sensing image; ground object segmentation; deep learning; feature fusion; loss function

收稿日期：2021-11-25

修回日期：2022-05-19

基金项目：安徽省能源互联网联合重点基金资助项目（2008085UD03）