无人机高光谱遥感估算冬小麦叶面积指数

为探讨利用低空无人机平台和高光谱影像对冬小麦叶面积指数进行遥感估算，该研究以拔节期冬小麦小区试验为基础，对原始冠层光谱进行一阶导数和连续统去除光谱变换，并在此基础上提取任意两波段组合的差值光谱指数（Difference Spectral Index，DSI）、比值光谱指数（Ratio Spectral Index，RSI）和归一化光谱指数（Normalized Spectral Index，NDSI），以最优窄波段光谱指数进行叶面积指数估算模型的构建。结果表明，最优窄波段指数的构成波段主要位于红边区域，最优窄波段指数与叶面积指数均呈现非线性关系；光谱变换显著提升了光谱变量与叶面积指数的相关性，其中连续统去除光谱所获取的NDSI（738，822）光谱指数与叶面积指数的相关性最佳；窄波段光谱指数和随机森林回归算法的叶面积指数估算模型精度最高，其相对预测偏差为2.01，验证集的决定系数和均方根误差分别为0.77和0.27。基于随机森林回归算法的无人机高光谱叶面积指数估算模型能够准确地实现小区域的叶面积指数遥感填图，为后期作物长势、变量施肥等提供理论依据。

Estimation of winter wheat leaf area index based on UAV hyperspectral remote sensing

Abstract: Leaf Area Index (LAI) is closely related to crop transpiration, photosynthesis, and final yield. Fast, non-destructive, and accurate monitoring of the winter wheat LAI during the critical growth period is an important way to accurately grasp crop canopy structure, growth information, above-ground biomass, yield, and pests. In the past, most of the estimated data of LAI were inverted by ground spectrum, Unmanned Aerial Vehicle (UAV) multi-spectrum, and satellite multi-spectral remote sensing data. In this study, a low-altitude unmanned aerial vehicle platform and an imaging spectrometer were used to obtain hyperspectral remote sensing images of jointing winter wheat in Xianyang city, Shaanxi province, China. The first derivative and continuous spectrum removal spectrum were used to transform the original canopy spectrum at 450-950 nm to extract any two frequency bands Specific. The Ratio Spectral Index (RSI), Differential Spectral Index (DSI), and Normalized Spectral Index (NDSI) were constructed respectively. Combining a narrow-band spectral index, a univariate regression analysis was performed on the best nine narrow-band spectral indices of a class, the original canopy spectrum, the first derivative spectrum, and the three best combinations spectra of any two bands under the continuous spectrum removal spectrum. The index was modeled using multiple linear regression. Two machine learning algorithms (BP neural network and random forest) were used to model a total of nine best narrowband spectral indices under three transformations. By comparing the coefficient of determination, the root mean square error, and the residual prediction deviation, the best estimation model of winter LAI was selected. The results showed that spectral transformation significantly improved the correlation between spectral variables and LAI, and the position with a higher correlation was mainly located in the red edge area. The correlation between the optimal narrow-band spectral index extracted based on the first derivative transform spectrum and the continuous spectrum removal transform spectrum and LAI was nonlinear, which was more suitable for fitting a quadratic function of a variable. Compared with the unary regression model, the accuracy of the multiple linear regression model based on the multi-spectral index was not significantly improved. However, both the multiple regression model and the unary linear regression model showed that the accuracy of the model based on the continuous spectrum transformation spectrum was higher than the accuracy of the model under the original spectrum and the first derivative transformation spectrum. It showed that the continuum removal method was used to transform the spectrum and used for modeling, and the NDSI (738,822) index based on the continuous removal spectrum had a good LAI estimation ability. Compared with the traditional regression model, the accuracy of the BP neural network model and the random forest regression model constructed with the 9 best narrowband spectral indices as independent variables had been significantly improved. Among them, the random forest regression model had the highest accuracy, because the random forest algorithm modeling could well tolerate some noise and outliers. As long as the parameters were adjusted accurately, overfitting was less likely to occur, and it was more suitable to solve certain nonlinear problems. Based on this model, the spatial distribution of the LAI of winter wheat at the jointing stage in this study area was basically in line with the actual situation, and the LAI estimation accuracy based on the narrow-band spectral index, and random forest algorithm was the highest (the residual prediction deviation was 2.01, the coefficient of determination was 0.77, the root mean square error was 0.27), which could be used as a basic model for hyperspectral remote sensing estimation of winter wheat UAV at jointing stage. The basic model could realize the LAI remote sensing mapping of a small area. It could provide a theoretical basis for later crop growth and variable fertilization.