基于改进AdvSemiSeg的半监督遥感影像作物制图方法  

Semi-supervised Network for Remote Sensing Crop Mapping Based on Improved AdvSemiSeg

作　　者：翟雪东 韩文霆[1,2] 马伟童 崔欣 李广 黄沈锦 ZHAI Xuedong;HAN Wenting;MA Weitong;CUI Xini;LI Guang;HUANG Shenjin(College of Mechanical and Electronic Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;Institute of Water-saving Agriculture in Arid Areas of China,Northwest A&F University,Yangling,Shaanxi 712100,China;College of Water Resources and Architectural Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;Chongqing Institute of Green and Intelligent Technology,Chinese Academy of Sciences,Chongqing 400714,China;Faculty of Computing,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]西北农林科技大学机械与电子工程学院,陕西杨凌712100 [2]西北农林科技大学中国旱区节水农业研究院,陕西杨凌712100 [3]西北农林科技大学水利与建筑工程学院,陕西杨凌712100 [4]中国科学院重庆绿色智能技术研究院,重庆400714 [5]哈尔滨工业大学计算学部,哈尔滨150001

出　　处：《农业机械学报》2024年第8期196-204,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2022YFD1900802);国家自然科学基金项目(51979233);中央高校基本科研业务费专项资金项目(2452023078)。

摘　　要：作物精准遥感制图对于农业资源调查与管理具有重要意义。深度学习为实现精准高效作物制图提供了技术支持。为了缓解深度学习对标记样本的依赖,本文提出了一种改进AdvSemiSeg的半监督遥感影像作物制图方法。所提方法引入STMF-DeepLabv3+作为对抗学习中的生成网络,通过Swin Transformer(ST)和多尺度特征融合(Multi-scale fusion, MF)模块提高生成网络特征编码能力和语义表达能力,改善遥感影像作物分割效果;此外,在判别网络中引入通道注意力(Efficient channel attention, ECA)模块,对不同通道特征图的表征信息进行自适应学习,增强判别网络对不同通道特征的感知能力。模型训练过程中,判别网络为生成网络提供高质量的伪标签和对抗损失,有效提高生成网络的泛化能力。采用所提方法与几种先进的半监督语义分割方法对内蒙古河套灌区遥感影像种植信息进行提取,本文方法性能最优。Crop precision remote sensing mapping holds significant importance for agricultural resource surveys and management.Deep learning provides technical support for achieving accurate and efficient crop mapping.To alleviate the dependency of deep learning on labeled samples,an improved semi-supervised remote sensing crop mapping method was proposed based on AdvSemiSeg.The proposed method introduced STMF-DeepLabv3+as the generator in the adversarial learning framework,enhancing the feature encoding and semantic expression capabilities of the generator through Swin Transformer(ST)and multi-scale fusion(MF)modules,thus improving the segmentation performance of remote sensing crop images.Additionally,the efficient channel attention(ECA)module was introduced after each convolutional layer of the discriminator to adaptively learn the representation information of different channel feature maps,enhancing the discriminator's perception of different channel features.During the training process,the discriminator provided high-quality pseudo-labels and adversarial losses to the generator,effectively improving the generalization ability of the generator.Compared with several advanced semi-supervised semantic segmentation methods,the proposed method achieved optimal performance in extracting planting information from remote sensing images in the Hetao Irrigation District of Inner Mongolia.

关 键 词：遥感 作物制图 半监督学习 生成对抗网络 多尺度特征融合 通道注意力 

分 类 号：S127[农业科学—农业基础科学]