地块尺度农作物遥感分类及其不确定性分析  

Remote sensing-based classification of crops on a farmland parcel scale and uncertainty analysis

作　　者：张冬韵 吴田军[2] 李曼嘉 郭逸飞 骆剑承 董文 ZHANG Dongyun;WU Tianjun;LI Manjia;GUO Yifei;LUO Jiancheng;DONG Wen(State Key Laboratory of Remote Sensing Science,Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100101,China;School of Land Engineering,Chang’an University,Xi’an 710064,China;College of Resources and Environment,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]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 [2]长安大学土地工程学院,西安710064 [3]中国科学院大学资源与环境学院,北京100049

出　　处：《自然资源遥感》2024年第4期124-134,共11页Remote Sensing for Natural Resources

基　　金：国家重点研发计划项目“地理空间智能核心技术与软件系统”(编号:2021YFB3900905)资助。

摘　　要：利用遥感技术开展农作物空间分布的快速调查与精准制图是现代精细农业的一项基础工作。然而,由于遥感影像获取、处理与分析过程中的局限性,传统农作物种植结构遥感制图精度受到一定影响,亟须面向农作物分类开展不确定性的空间建模与特征分析。该文将耕地地块作为基本空间单元,选择宁夏引黄灌区作为试验区,利用多源遥感数据和机器学习算法实现地块尺度的农作物分类,进而基于混合熵构建不确定性计算模型,生成地块农作物类型不确定性空间分布,再利用多源辅助数据对地块农作物分类不确定性进行回归建模,探究相关地理变量的潜在影响。实验结果表明:①耕地提取及分类环节共构建地块矢量单元149万个,总体作物分类精度达0.80,制图结果与实际农业耕作管理单元相匹配,分类效果较之传统的像素尺度方法更为理想;②地块尺度的农作物分类不确定性总体较低,存在较为显著的类别差异,水稻、菜地、苜蓿分类不确定性较小,单种、复种小麦不确定性较高,玉米作物分类不确定性介于二类之间;③地块级作物分类不确定性与种植结构、资源条件等多种环境因素有关,且与作物类型、水源可达性的相关性最为显著。The rapid survey and accurate mapping of the spatial distribution of crops using remote sensing are fundamental to modern precision agriculture.However,limitations in the acquisition,processing,and analysis of remote sensing images impact the mapping accuracy of traditional crop planting structures.Therefore,there is an urgent need to conduct spatial modeling and feature analysis for the uncertainty in crop classification.Using the Ningxia Yellow River irrigation area as a trial area and farmland parcels as the basic spatial units,this study classified crops on a parcel scale utilizing multi-source remote sensing data and machine learning algorithms.Then,an uncertainty calculation model was constructed based on mixed entropy,yielding the spatial distribution of the uncertainty of crop types in farmland parcels.Afterward,multi-source auxiliary data were employed to build a regression model for the uncertainty,and the potential impacts of related geographical variables on the uncertainty were explored.The experiment results indicate that 1.49 million vector units were constructed for the farmland parcels during the farmland extraction and classification session,yielding an overall crop classification accuracy of 0.80.The mapping results aligned well with the actual agricultural management units,and the classification results proved more better than the traditional pixel-based methods.The uncertainty in the parcel-scale crop classification was generally lower,with significant differences among crop types.The uncertainty was low for rice,vegetable plots,and alfalfa,relatively higher for wheat of single-and double-cropping patterns,and moderate for maize.The uncertainty in parcel-scale crop classification is influenced by various environmental factors such as planting structure and resource conditions,exhibiting the most significant correlations with crop type and water accessibility.

关 键 词：遥感 农作物分类 地块 不确定性 机器学习 混合熵 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]