耦合多源无人机遥感数据和机器学习方法的玉米SPAD估测  

Maize SPAD estimation by combining multi-source unmanned aerial vehicle remote sensing data and machine learning methods

作　　者：周科 陈鹏飞[1,3] ZHOU Ke;CHEN Peng-Fei(Institute of Geographic Sciences and Natural Resources Research,Chinese Academy of Sciences/State Key Laboratory of Resources and Environment Information System,Beijing 100101,China;University of Chinese Academy of Sciences,Beijing 100049,China;Jiangsu Center for Collaborative Innovation in Geographic Information Resource Development and Application,Nanjing 210023,Jiangsu,China)

机构地区：[1]中国科学院地理科学与资源研究所/资源与环境信息系统国家重点实验室,北京100101 [2]中国科学院大学,北京100049 [3]江苏省地理信息资源开发与利用协同创新中心,江苏南京210023

出　　处：《作物学报》2025年第5期1389-1399,共11页Acta Agronomica Sinica

基　　金：中国科学院先导A专项(XDA28040502);国家自然科学基金项目(41871344)资助。

摘　　要：为实现玉米精准施肥管理,准确识别其叶绿素含量具有重要意义。叶片叶绿素相对含量(soil and plant analyzer development,SPAD)值是叶绿素含量的重要指示参数,已有研究多采用单一数据源,结合单一机器学习方法来对其反演。为提高SPAD反演精度,本研究探讨耦合多源无人机遥感影像与多种机器学习方法来开展SPAD值反演的可行性,并将其与已有方法进行比较。基于不同有机肥、无机肥、秸秆还田以及种植密度处理的玉米田间试验,获取玉米四叶期和九叶期的无人机多光谱影像和RGB(red-green-blue)影像,并同步测量了叶片SPAD数据。基于多尺度分析的方法,将RGB影像与多光谱影像进行融合,生成既具有高空间分辨率又具有多光谱的融合影像。此外,基于集成学习思想,选择BP-人工神经网络法(back propagation-artificial neural network,BP-ANN)、支持向量机法(support vector machine,SVM)、广义加性模型法(generalized additive model,GAM)、随机森林法(random forest,RF)等不同类型机器学习模型,构建集成学习模型(ensemble learning method,ELM)。基于以上数据源和模型,设计不同数据源和不同机器学习模型的耦合情景。将数据集分为建模数据集和验证数据集,基于建模数据集构建每种情景下的SPAD反演模型,然后基于验证数据集进行模型验证,对比分析确定最优SPAD反演模型与数据源。相对于单源数据,多源数据通过融合多光谱影像的光谱信息和RGB影像的纹理信息,提高了SPAD反演的精度。此外,相对于单一机器学习方法,基于集成学习思想耦合多种机器学习方法可以提高SPAD的反演精度。在所有情景中,基于ELM方法和融合影像的SPAD模型精度最高,其建模R_(cal)^(2)为0.83、RMSEcal为1.93,验证R_(val)^(2)为0.80、RMSEval为2.07;其他情景下,各模型的建模R_(cal)^(2)在0.64~0.88之间,RMSEcal在1.63~2.84之间;验证R_(val)^(2)在0.60~0.78之间,RMSEval在2.18~3.01Accurately identifying chlorophyll content is essential for precise fertilization management in maize.The SPAD(Soil and plant analyzer development)value of leaves serves as a reliable indicator of chlorophyll content.For SPAD prediction using remote sensing,most existing studies rely on single data sources combined with machine learning methods.To enhance SPAD prediction accuracy,this study explores the feasibility of integrating multi-source unmanned aerial vehicle(UAV)data with various machine learning methods,comparing the results to traditional approaches.A maize field experiment was conducted with different treatments,including organic fertilizer,inorganic fertilizer,straw return,and varying planting densities.UAV multispectral and RGB images were acquired at the V4 and V9 growth stages,and SPAD values of maize leaves were measured subsequently.Using a multi-scale analysis approach,RGB images were fused with multispectral images to produce a dataset combining high spatial resolution with multispectral information.Additionally,an ensemble learning method(ELM)was developed by integrating multiple machine learning models,including the backpropagation artificial neural network(BP-ANN),support vector machine(SVM),generalized additive model(GAM),and random forest(RF).Different scenarios were designed by coupling various data sources and machine learning models.The dataset was divided into calibration and validation subsets.SPAD prediction models were developed by calibration dataset,and their performance was evaluated using the validation dataset.Comparative analysis identified the optimal model and data source.Results showed that multi-source data significantly improved SPAD prediction accuracy by combining the spectral information of multispectral images with the texture information of RGB images.Furthermore,the ensemble learning method outperformed single machine learning methods,achieving higher SPAD prediction accuracy.Among all scenarios,the SPAD prediction model using the ELM method and fused images exhibit

关 键 词：机器学习 多源数据 玉米 SPAD 无人机 

分 类 号：S127[农业科学—农业基础科学]