基于多光谱影像和机器学习算法的红树林树种LAI估算

为了提高无人机遥感对冬小麦叶面积指数（leaf area index，LAI）反演模型的精度与泛化能力，该研究利用无人机搭载多光谱相机获取不同氮素处理和不同复种方式的冬小麦生长实测数据，结合PROSAIL辐射传输模型生成包含机理信息的模拟数据，基于不同组合方式建立了5种LAI反演混合数据集，结合多种机器学习方法，以期构建经验与机理相结合的LAI高精度反演模型。由于LAI反演受近红外波段（near infrared，NIR）反射率影响大，该研究筛选7种与NIR波段相关的植被指数提取冬小麦光谱特征，构建与混合数据集LAI的相关系数矩阵，进一步探究不同光谱特征对冬小麦LAI的影响程度。在此基础上，采用具有代表性和普适性的4种机器学习方法，即贝叶斯岭回归模型、线性回归模型、弹性网络模型和支持向量回归模型，构建不同冬小麦LAI反演模型，用以评估基于半经验半机理数据反演冬小麦LAI的可行性，进一步探索其对不同氮素水平和复种方式的冬小麦长势评估能力。结果表明：1）筛选的与NIR波段相关的植被指数与冬小麦LAI之间存在较强的相关性，其中归一化差异植被指数、增强植被指数、归一化差异红边指数、比值植被指数、红边叶绿素植被指数、土壤调节植被指数与LAI呈正相关，结构不敏感色素植被指数与LAI呈负相关；2）辐射传输模型中体现了冬小麦LAI影响太阳光线传播的机理，结果表明，与实测数据混合建立的模型，具有较强的鲁棒性和泛化能力。相比于其他3种模型，支持向量回归模型在各种数据组合下均取得了较好的LAI预测性能，在C1、C2、C3、C4这4种训练-测试组合的训练集中R²依次为0.86、0.87、0.88、0.91，RMSE依次为0.47、0.45、0.45、0.41；在测试集的R²依次为0.85、0.19、0.89、0.87，RMSE依次为0.45、1.31、0.49、0.50；3）使用支持向量机生成试验区LAI反演图，对4种氮素水平和2种复种方式的冬小麦长势评估，结果表明，适当的施加氮素处理能提高冬小麦LAI值，麦-豆复种方式下的冬小麦LAI值普遍高于麦-玉复种的LAI值。该研究为冬小麦LAI的反演提供了一种有效的方法，并为高效评估冬小麦长势研究提供了参考。

Comparing different methods for wheat LAI inversion based on hyperspectral data

Leaf area index (LAI) is one of the key indicators in the structure and function of vegetation canopy, in order to estimate the biomass and crop growth. This study aims to improve the accuracy and generalization of the LAI inversion model for the winter wheat using unmanned aerial vehicle (UAV) remote sensing. An inversion model was established using semi-empirical and semi-mechanistic approaches. An UAV with a multispectral camera was utilized to obtain the measured data of winter wheat growth with different nitrogen treatments and replanting. PROSAIL radiative transfer model was used to generate the simulated data with mechanistic information. Five LAI inversion hybrid datasets were established using different combinations of measured and simulated data. Various machine learning methods were used to construct a high-precision LAI inversion model using empirical and mechanistic information. Seven kinds of vegetation indices related to NIR bands were screened to extract the winter wheat spectral features, in order to reduce the reflectance of NIR bands. The correlation coefficient matrix between the vegetation indices and the LAI of the mixed dataset was calculated to further explore the degree of influence of different spectral features on the LAI of winter wheat. The LAI inversion models of winter wheat were formed using Bayesian ridge regression, linear regression, elasticity network, and support vector regression model. The feasibility of LAI inversion was also evaluated using semi-empirical and semi-mechanistic data. The ability of the improved model was finally determined to assess the winter wheat growth for different nitrogen levels and replanting. The results showed that: 1) There was a strong correlation between the screened vegetation indices associated with NIR bands and winter wheat LAI. Normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), normalized difference red edge index (NDRE), ratio vegetation index (SR), red edge chlorophyll vegetation index (RECI), and soil adjusted vegetation index (SAVI) were positively correlated with the LAI, whereas, the structurally insensitive pigment vegetation index (SIPI) was negatively correlated with the LAI. 2) The radiative transfer model was represented for the winter wheat LAI subjected to the propagation of solar rays. The strong robustness and generalization were achieved to mix with the measured data. The support vector regression (SVR) model achieved better LAI prediction performance under various data combinations, compared with the rest. In the training set of the four training-test combinations C1, C2, C3 and C4, R² is 0.86, 0.87, 0.88, 0.91, RMSE is 0.47, 0.45, 0.45, 0.41; in the test set, R² is 0.85, 0.19, 0.89, 0.87, RMSE is 0.45, 1.31, 0.49, 0.50. 3) A support vector machine model was used to generate the LAI inversion maps for the test area. The winter wheat growth was evaluated under four nitrogen levels and two replanting models. The results showed that 180 kg/hm² fertilization was more effective than 135 kg/hm² one, but 225 kg/hm² fertilization was similar to 180 kg/hm² one. An optimal application of nitrogen treatment can be expected to improve the LAI value of winter wheat. Among them, the LAI values under wheat-bean replanting were generally higher than those of wheat-yue replanting. This finding can provide an effective way for the inversion of winter wheat LAI in the efficient assessment of winter wheat growt