Decision tree inversion model results.

As a key substance for crop photosynthesis, chlorophyll content is closely related to crop growth and health. Inversion of chlorophyll content using unmanned aerial vehicle (UAV) visible light images can provide a theoretical basis for crop growth monitoring and health diagnosis. We used rice at the tasseling stage as the research object and obtained UAV visible orthophotos of two experimental fields planted manually (experimental area A) and mechanically (experimental area B), respectively. We constructed 14 vegetation indices and 15 texture features and utilized the correlation coefficient method to analyze them comprehensively. Then, four vegetation indices and four texture features were selected from them as feature variables to be added into three models, namely, K-neighborhood (KNN), decision tree (DT), and AdaBoost, respectively, for inverting chlorophyll content in experimental areas A and B. In the KNN model, the inversion model built with BGRI as the independent variable in region A has the highest accuracy, with R2 of 0.666 and RSME of 0.79; the inversion model built with RGRI as the independent variable in region B has the highest accuracy, with R2 of 0.729 and RSME of 0.626. In the DT model, the inversion model built with B-variance as the independent variable in region A has the highest accuracy, with R2 of 0.840 and RSME of 0.464; the inversion model built with G-mean as the independent variable in region B has the highest accuracy, with R2 of 0.845 and RSME of 0.530. In the AdaBoost model, the inversion model built with R-skewness as the independent variable in region A has the highest accuracy, with R2 of 0.826 and RSME of 0.642; the inversion model established with g as the independent variable in area B had the highest accuracy, with R2 of 0.879 and RSME of 0.599. In the comprehensive analysis, the best inversion models for experimental areas A and B were B-variance-decision tree and g-AdaBoost, respectively, whose models can quickly and accurately carry out the inversion of chlorophyll content of rice, and provide a theoretical basis for the monitoring of the crop’s growth and health under different cultivation methods.

叶绿素作为作物光合作用的关键物质，其含量与作物生长态势及健康状况密切相关。利用无人机（unmanned aerial vehicle, UAV）可见光影像反演叶绿素含量，可为作物生长监测与健康诊断提供理论依据。本研究以抽穗期水稻为研究对象，分别获取了人工种植（试验区A）与机械种植（试验区B）两块试验田的无人机可见光正射影像。本研究构建了14种植被指数与15种纹理特征，并采用相关系数法对其进行综合分析。随后从中筛选出4种植被指数与4种纹理特征作为特征变量，分别代入K近邻（K-neighborhood, KNN）、决策树（decision tree, DT）以及AdaBoost三种模型中，对试验区A、B的叶绿素含量进行反演。在K近邻模型中，试验区A以BGRI为自变量构建的反演模型精度最高，决定系数（R²）为0.666，均方根误差（RSME）为0.79；试验区B以RGRI为自变量构建的反演模型精度最高，决定系数（R²）为0.729，均方根误差（RSME）为0.626。在决策树模型中，试验区A以B-variance为自变量构建的反演模型精度最高，决定系数（R²）为0.840，均方根误差（RSME）为0.464；试验区B以G-mean为自变量构建的反演模型精度最高，决定系数（R²）为0.845，均方根误差（RSME）为0.530。在AdaBoost模型中，试验区A以R-skewness为自变量构建的反演模型精度最高，决定系数（R²）为0.826，均方根误差（RSME）为0.642；试验区B以g为自变量构建的反演模型精度最高，决定系数（R²）为0.879，均方根误差（RSME）为0.599。综合分析结果表明，试验区A与B的最优反演模型分别为B-variance-决策树模型与g-AdaBoost模型，二者可快速精准地反演水稻叶绿素含量，为不同种植模式下的作物生长与健康监测提供理论依据。