耦合NDVI与纹理时序特征的地块作物遥感分类  

Land parcel crop remote sensing classification via coupleing with time series features of NDVI and texture

作　　者：史洁宁 吴田军[2] 黄启厅 骆剑承 任应超 徐欣雨 SHI Jie-ning;WU Tian-jun;HUANG Qi-ting;LUO Jian-cheng;REN Ying-chao;XU Xin-yu(School of Sciences,Chang’an University,Xi’an,Shaanxi 710064,China;School of Land Engineering,Chang’an University,Xi’an,Shaanxi 710064,China;Agricultural Science and Technology Information Research Institute,Guangxi Academy of Agricultural Sciences,Nanning,Guangxi 530007,China;State Key Laboratory of Remote Sensing Science,Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100101,China;College of Resources and Environment,University of Chinese Academy of Sciences,Beijing 100049,China;National Engineering Research Center for Geomatics,Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China)

机构地区：[1]长安大学理学院,陕西西安710064 [2]长安大学土地工程学院,陕西西安710064 [3]广西农业科学院农业科技信息研究所,广西南宁530007 [4]中国科学院空天信息创新研究院遥感科学国家重点实验室,北京100101 [5]中国科学院大学资源与环境学院,北京100049 [6]中国科学院空天信息创新研究院国家遥感应用工程技术研究中心,北京100094

出　　处：《南方农业学报》2025年第1期29-40,共12页Journal of Southern Agriculture

基　　金：国家重点研发计划项目(2021YFB3900905);河北省中央引导地方科技发展资金项目(236Z0104G)。

摘　　要：【目的】充分挖掘遥感影像的时间和空间信息,准确识别地块作物类型,为作物类型空间分布制图、产量估计及农业生产决策等提供可靠的数据支持。【方法】以Google Earth影像为参考,获得美国加利福尼亚州金斯县完整的地块边界,利用多时相Sentinel-2影像构建地块归一化植被指数(NDVI)时间序列和时间—纹理二维表征图作为分类特征,NDVI时间序列捕捉作物生长的物候变化,时间—纹理二维表征图捕捉空间特征随时间的动态变化,进而使用卷积神经网络(CNN)+长短时记忆网络(LSTM)双流架构来联合时间和空间特征实现农田作物的准确识别。【结果】与仅使用NDVI时序的传统方法相比,纳入纹理时序后的方法明显提高分类精度,随机森林的分类精度由0.89提升至0.93,支持向量机的分类精度由0.88提升至0.93,表明加入空间特征的纹理时序能有效提升作物分类能力;而使用CNN+LSTM双流架构分类模型进行地块作物分类的总体精度达0.95,特别是葡萄和冬小麦的分类精度提升效果明显,F_(1)分别提升至0.90和0.92,表明相较于传统的分类器,使用CNN+LSTM双流架构可实现更精准的地块作物识别。【建议】在种植结构复杂、农作物生长习性相近的地区进行地块作物遥感分类时,考虑将纹理时序特征纳入分类体系,并使用CNN+LSTM双流架构分别捕捉作物生长的时间和空间特征。这种综合应用时间和空间信息的方法,能提升地块作物分类的准确度。【Objective】To fully explore the temporal and spatial information of remote sensing images,accurately identified crop types on land parcels,which could provide reliable data support for crop type spatial distribution mapping,yield estimation and agricultural production decision-making,etc.【Method】Using Google Earth images as a reference,obtained the complete boundary of the study area of Kings County,California,United States,and used multi-temporal Sentinel-2 images to construct NDVI time series and time-texture two-dimensional representation maps as classification features.The NDVI time series captured the phenological changes of crop growth,while the time-texture two-dimensional representation maps captured the dynamic changes of spatial features over time.Then,convolutional neural network(CNN)+long short-term memory(LSTM)dual-stream architecture was used to combine temporal and spatial features to achieve accurate recognition of crops.【Result】The experimental results showed that compared with traditional methods that only used NDVI time series,the method incorporating texture time series greatly improved classification accuracy.The classification accuracy of the random forest increased from 0.89 to 0.93,and the classification accuracy of the sup‐port vector machine increased from 0.88 to 0.93.This indicated that the texture time series with spatial features effectively improved crop classification ability.The overall accuracy of using the CNN+LSTM dual-stream architecture classification model for land parcel crop classification reached 0.95,in particular,the classification accuracy of grape and winter wheat improved greatly,F_(1) increased to 0.90 and 0.92 respectively.This demonstrated that,compared to traditional classifiers,the CNN+LSTM dual-stream architecture achieved more accurate land parcel crop recognition.【Suggestion】When con‐ducting remote sensing classification of land parcels crops in areas with complex planting structures and similar crop growth habits,it is considered to incor

关 键 词：作物分布 地块尺度 归一化植被指数(NDVI) 时间序列 空间纹理特征 CNN+LSTM双流架构 

分 类 号：S127[农业科学—农业基础科学]