基于时序深度学习模型的河套灌区作物分类研究  

Crop Type Mapping in the Hetao Irrigation District Based on a Temporal Deep Learning Model

作　　者：杜帅 肖玲君 查元源 刁雨晴 连勰 DU Shuai;XIAO Ling-jun;ZHA Yuan-yuan;DIAO Yu-qing;LIAN Xie(School of Water Resources and Hydropower Engineering,Wuhan University,Wuhan 430072,Hubei Province,China;Wuhan Yuchi Testing Technical Co.,Ltd.,Wuhan 430072,Hubei Province,China;Hubei Water Resources Research Institute,Wuhan 430070,Hubei Province,China;Yellow River Engineering Consulting Co.,Ltd,Zhengzhou 450003,Henan Province,China)

机构地区：[1]武汉大学水利水电学院,湖北武汉430072 [2]武汉市宇驰检测技术有限公司,湖北武汉430072 [3]湖北省水利水电科学研究院,湖北武汉430070 [4]黄河勘测规划设计研究院有限公司,河南郑州450003

出　　处：《节水灌溉》2025年第4期1-7,15,共8页Water Saving Irrigation

基　　金：国家重点研发计划(2021YFC3201204);国家自然科学基金资助项目(52279042)。

摘　　要：及时、准确地绘制作物分类地图是智慧灌区建设的重要基础,可为农作物生长监测、产量预测及农业生态环境评估提供关键数据支撑。深度学习模型在作物分类研究中表现出卓越的效果,但在处理时序遥感数据时仍存在一定局限性。基于Google Earth Pro软件、GEE平台和geemap工具,生成了2020年河套灌区的作物分类地图,并对其精度和可靠性进行了评估。以该分类结果为标签,利用Sentinel-2图像生成了包含河套灌区5-10月72个波段的Sentinel-2数据集和6个波段的EVI数据集,比较了多种机器学习模型与深度学习模型在河套灌区的作物分类性能。研究结果表明,基于时序的深度学习方法(TFBS)取得了最佳分类精度,其mIoU、mprecision和mrecall分别达到0.8722、0.9247和0.9260。在处理72个波段遥感影像数据集和EVI数据集时,基于时序的深度学习模型展现出较强的鲁棒性,而Unet模型在处理EVI数据时无法收敛,难以提取时序特征。研究表明:基于时序的深度学习模型具备更高的分类精度和显著的鲁棒性,为作物分类研究中的模型选择提供了参考。其分类结果也为河套灌区的农业管理提供了重要的技术支持与数据保障。Accurate and timely crop type mapping is essential for smart irrigation district development,providing crucial data for crop growth monitoring,yield estimation,and agricultural ecological assessments.Deep learning models have demonstrated superior performance in crop classification research,offering great potential for automating and enhancing classification accuracy.However,deep learning models still face limitations in handling time-series remote sensing data,requiring further refinement.This study focuses on crop classification in the Hetao Irrigation District using remote sensing data from Sentinel-2 satellites.To address time-series data challenges,linear interpolation was used to synthesize data from May to October 2020.The Enhanced Vegetation Index(EVI)was derived from the imagery to create a time-series dataset tracking vegetation changes throughout the growing season.Various deep learning models were used to analyze crop classification within the Hetao Irrigation District,comparing the performance of different approaches.Among the tested methods,the Temporal Featurebased Segmentation(TFBS)model achieved the highest classification accuracy,with mean Intersection over Union(mIoU),mean precision,and mean recall values of 0.8722,0.9247,and 0.9260,respectively.These metrics indicate that the TFBS model outperformed other models.The TFBS model proved robust in handling both raw remote sensing imagery and EVI data,effectively extracting temporal features for improved classification accuracy.In contrast,the Unet model failed to converge with EVI data,struggling to capture time-series characteristics.This suggests that while Unet can be effective for certain types of image classification tasks,it struggles with the complexity of temporal data derived from vegetation indices,making it less suitable for time-series remote sensing applications in crop classification.The success of the TFBS model highlights the importance of selecting models designed for time-series data when working with dynamic factors like crop gr

关 键 词：作物分类 深度学习 河套灌区 遥感 时序 

分 类 号：S27[农业科学—农业水土工程]