基于三波段光谱指数的春小麦叶片水分含量估算

为探讨利用三波段植被指数（three-band index, 3BI）对春小麦叶片水分含量（leaf water content, LWC）估算的可行性，在田间尺度上，利用ASD-FieldSpec-3光谱仪测定春小麦抽穗期冠层光谱反射率，采用任意波段组合方法，分别建立两波段植被指数（two-band index, 2BI）包括比值植被指数（RVI）、归一化植被指数（NDVI）、差值植被指数（DVI）及3BI，并对单波段反射率、两波段植被指数和三波段植被指数与春小麦抽穗期LWC之间进行相关性分析，筛选稳定的光谱参数，基于人工神经网络（artificial neural network, ANN）、K近邻（K-nearest neighbors, KNN）和支持向量回归（support vector regression, SVR）等3种机器学习算法，建立有效波段组合运算的抽穗期春小麦LWC估算模型，并利用独立样本对模型精度进行检验和评价。结果表明，单波段反射率、2BI和3BI与春小麦抽穗期LWC之间的相关性均达极显著水平（P＜0.01），而相关系数差异较大，绝对值分别为0.23、0.62、0.94，说明组合波段展现了光谱隐含信息，避免有效光谱信息的丢失；估算模型中，春小麦抽穗期以KNN算法和最佳3BI组合变量（3BI-5_{(1075, 1095, 1085)}、3BI-6_{(1100, 400, 1097)}）构建的模型拟合度最高（r²=0.83），均方根误差最小（RMSE=2.14%），相对偏差百分比超出了2.0以上（RPD=2.31），说明该模型具有一定的预测能力。由此可见，通过任意波段组合，可明显提高3BI与春小麦LWC的关联度，且基于K近邻算法构建的模型具有较好的稳定性和估算能力。

Estimation of Leaf Water Content of Spring Wheat Based on 3D Spectral Index

To explore the feasibility of using three band vegetation index (3BI) to estimate leaf water content (LWC) of spring wheat, the ASD-FieldSpec-3 spectrometer was used at the field scale to measure the canopy spectral reflectance of spring wheat at heading stage. Combination of wave bands was used to establish two band vegetation index (2BI), including ratio vegetation index (RVI) and normalized vegetation index (NDVI), and Difference Vegetation Index (DVI) and 3BI were used to analyze the correlation between single band reflectance, two band vegetation index, and three band vegetation index with LWC at heading stage of spring wheat. Stable spectral parameters were selected, and three machine learning algorithms, including artificial neural network (ANN), K-nearest neighbors (KNN), and support vector regression (SVR) were conducted to establish an effective band combination operation based LWC estimation model for spring wheat at heading stage, and to test and evaluate the accuracy of the model using independent samples. The results showed that the correlation between single band reflectance, 2BI and 3BI, and LWC at heading stage of spring wheat reached a highly significant level (P<0.01), with significant differences in correlation coefficients, with absolute values of 0.23, 0.62, and 0.94, respectively. This indicates that the combined bands exhibit spectral implicit information and avoid the loss of effective spectral information. For the estimation model, the model built with KNN algorithm and the best 3BI combination variables (3BI-5_{(1075, 1095, 1085)}, 3BI-6_{(1100, 400, 1097)} at the heading stage of spring wheat had the highest fitting degree (r²=0.83), the smallest Root-mean-square deviation (RMSE=2.14%), and the relative deviation percentage more than 2.0 (RPD=2.31), indicating that the model had certain predictive ability. It is suggested that the correlation between 3BI and spring wheat LWC is significantly improved through the combination of wave bands. The model constructed based on the K-nearest neighbor algorithm has good stability and estimation ability.