摘要
针对当前耕地“非粮化”监测任务重、传统自动提取方法应用于耕地“非粮化”业务难度大的问题,该文以广西壮族自治区为研究区域,研究深度学习遥感变化检测应用于耕地“非粮化”变化自动提取的可行性,提出基于余弦相似度的差异注意力模块以增强变化内容的特征表示,面向业务应用权衡不同分类阈值下召回率与精确率的优劣。经实验验证,该文模型的召回率、精确率分别约为78%与73%;相较基线方法,该文差异注意力模块可提升约3%的精度;在业务应用中,选择适当的模型分类阈值能够以仅25%的额外核查工作换取1.41%的召回率提升。实验结果表明了深度学习遥感变化检测用于耕地“非粮化”自动提取业务的可行性,为大范围耕地“非粮化”监测的业务应用提供了参考。
In view of the problem that it is difficult to apply traditional methods to automatic extraction of non-grain conversion,taking Guangxi Zhuang Autonomous Region as the research area,the feasibility of the non-grain conversion extraction with deep learning change detection was explored in this paper.A difference attention module based on cosine similarity was proposed to enhance the feature representation of changing content.In addition,the trade-off between recall and accuracy under different thresholds of model was considered for practical applications.Through experimental verification,the change detection model could effectively extract non-grain conversions with about 78%recall and 73%precision.The proposed difference attention module improved recall and precision by about 3%compared with the baseline method.By selecting an appropriate classification threshold,the proposed model traded off 25%extra polygon check against 1.41%higher recall.The result showed the feasibility of deep learning remote sensing change detection for non-grain conversion in cultivated land extraction,which provided a reference for the application of large-scale monitoring of non-grain conversion in cultivated.
作者
黄友菊
农志铣
韩广萍
吴慧
HUANG Youju;NONG Zhirian;HAN Guangping;WU Hui(Guangxi Zhuang Autonomous Region Institute of Natural Resources Remote Sensing,Nanning 530200,China)
出处
《测绘科学》
CSCD
北大核心
2023年第8期193-201,219,共10页
Science of Surveying and Mapping
基金
广西重点研发计划项目(桂科AB22080080)
高分辨率对地观测系统重大专项政府综合治理应用于规模化产业化示范项目(84-Y50G25-9001-22/23)。
关键词
非粮化
遥感
深度学习
变化检测
广西
non-grain conversion
remote sensing
deep learning
change detection
Guangxi
作者简介
黄友菊(1980-),女,广西南宁人,高级工程师,硕士,主要研究方向为遥感影像智能处理,E-mail:huangyouju2022@163.com;通信作者:农志铣,E-mail:ssgnzx@163.com。
