安徽农业大学学报2024，Vol.51Issue(5)：888-898,11.DOI:10.13610/j.cnki.1672-352x.20241101.002
基于Vision Transformer-LSTM(ViTL)的多时序遥感影像农作物分类方法
A crop classification method based on vision Transformer-LSTM(ViTL)for multi-temporal remote sensing images
摘要
Abstract
For the current research in remote sensing crop classification using deep learning models,there exists an inadequacy in sampling temporal and spatial information features,leads to challenges in accurately extracting crops due to issues such as boundary fuzziness,omissions,and misclassifications.In this study,we propose a deep learning model called Vision Transformer-long short term memory(ViTL)to addresses these challenges.The ViTL model integrated three key modules such as two-way Vit-Transformer feature extraction,spatio-temporal feature fusion,and long short-term memory(LSTM)temporal classification,and the two-way Vit-Transformer feature ex-traction module was used to capture the spatio-temporal feature correlation of the image,one way to extract spatial classification features and one way to extract temporal change features;spatio-temporal feature fusion module was used to cross-fertilize multi-temporal feature information,and the LSTM temporal classification module captures multi-temporal dependencies and performs output classification.In this study,the theory and method of remote sensing technology based on multi-temporal satellite images were comprehensively utilized to extract crop infor-mation in Nehe city,Qiqihar city,Heilongjiang Province.The results showed that,the ViTL model performed well,with its overall accuracy(OA),a mean intersection over union(MIoU)and F1 scores reached 0.867 6,0.698 7 and 0.8175,respectively,and the F1 scores of the ViTL model were improved by 9％-12％compared with other widely used deep learning methods,including 3-D convolutional neural networks(3-D CNNs),2-D convolutional neural networks(2-D CNNs),and long short-term memory recurrent neural networks(LSTMs),which shows significant superiority.The ViTL model provides a new idea for accurate and efficient crop classification.关键词
农作物分类/Vision Transformer(ViT)/LSTM/深度学习/遥感监测Key words
crop classification/vision transformer(ViT)/LSTM/deep learning/remote sensing monitoring分类
农业科技引用本文复制引用
张青云,杨辉,李兴伍,武永闯..基于Vision Transformer-LSTM(ViTL)的多时序遥感影像农作物分类方法[J].安徽农业大学学报,2024,51(5):888-898,11.基金项目
国家自然科学基金(42101381)和安徽省重点研发国际合作项目(202104b11020022)共同资助. （42101381）
