基于DE-Segformer的无人机影像农业大棚信息提取方法  

UAV images agricultural greehouse information extraction method based on DE-Segformer

作　　者：张琳翔 郭勇[1] 王浩宇 刘宏建[1] 王岩 ZHANG Linxiang;GUO Yong;WANG Haoyu;LIU Hongjian;WANG Yan(Information Engineering University,Zhengzhou 450052,China;Xinjiang Institute of Ecology and Geography,Chinese Academy of Sciences,Urumqi 830011,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]信息工程大学,郑州450052 [2]中国科学院新疆生态与地理研究所,乌鲁木齐830011 [3]中国科学院大学,北京100049

出　　处：《测绘工程》2024年第2期56-64,共9页Engineering of Surveying and Mapping

基　　金：国家自然科学基金资助项目(41971375)。

摘　　要：由于无人机影像具有空间分辨率高、机动性强、成本低等优点,在智慧农业中可以用来监测各种特征信息的变化。为了进一步提高农业大棚识别精度,提出基于Segformer解码器改进的DE-Segformer模型。其中,通过特征整合模块融合高层语义信息,实现对多尺度特征图隐式特征的深度挖掘;通过反卷积上采样模块实现对影像空间细节的准确还原表达。以潍坊-黄岛沿线为研究区开展实验,DE-Segformer的总体精度和交并比分别为0.979和0.957,与Segformer相比分别提高0.013和0.028,且锯齿问题得到明显缓解。消融实验证明设计的两个模块均起到积极作用,联合使用可达到最大精度。为实现无人机影像农业大棚的精细分割,以及基于Transformer的语义分割模型在遥感信息提取中的工程化应用提供参考。Unmanned aerial vehicle(UAV)imagery,with high spatial resolution,high mobility and low cost,are widely used for various information extraction and change monitoring in agricultural greenhouse.In order to further improve the recognition accuracy of agricultural greenhouse and expand the engineering application of deep learning technology in the field of remote sensing information extraction,an improved Segformer model,DE-Segformer,is proposed.A feature fusion module is designed to fuse high-level semantic information to achieve deep mining of implicit features.A deconvolution upsampling module is designed to achieve accurate reduction representation of spatial details.Comparative experiments were carried out along the Huangdao-Weifang route.The overall accuracy and intersection ratio of DE-Segformer were 0.979 and 0.957 respectively,which improved by 0.013 and 0.028 respectively compared with the Segformer.Besides,DE-Segformer significantly alleviates the jaggedness problem of Segformer and achieves fine segmentation of agricultural greenhouse.The ablation experiments further demonstrate that both modules play a positive role and that maximum accuracy can be achieved when used jointly.This paper provides a methodological reference for the engineering application of Transformer-based semantic segmentation models in remote sensing semantic segmentation.

关 键 词：区域产业格局 语义分割 深度学习 无人机影像 农业大棚 

分 类 号：P237[天文地球—摄影测量与遥感]