基于机器学习的棉花叶面积指数监测

为实现基于机器学习和无人机高光谱影像进行棉花全生育期叶面积指数(Leaf Area Index, LAI)监测,该研究基于大田种植滴灌棉花,在不同品种及不同施氮处理的小区试验基础上,对无人机获取的高光谱数据分别采用一阶导(First Derivative, FDR)、二阶导(Second Derivative, SDR)、SG(Savitzky-Golay)平滑和多元散射校正(Multiplicative Scatter Correction, MSC)进行预处理,并结合Pearson相关系数法、连续投影(Successive Projections Algorithm, SPA)、随机蛙跳(Shuffled Frog Leaping Algorithm, SFLA)和竞争性自适应重加权(Competitive Adaptive Reweighting, CARS)筛选敏感波段,将筛选出的波段,使用偏最小二乘回归(Partial Least Squares Regression, PLSR)、支持向量回归(Support Vector Regression, SVR)和随机森林回归(Random Forest Regression, RFR)3种机器学习算法构建棉花LAI监测模型。结果表明:棉花冠层LAI敏感响应波段集中在可见光(400～780 nm)和近红外(900 nm之后)波段;对比3种机器学习算法,各预处理下RFR建立的LAI监测模型精度最高,稳定性最好,其中以FDR-SFLA-RFR模型最佳,在建模集的决定系数为0.74,均方根误差为1.648 3,相对均方根误差为26.39%;验证集的决定系数、均方根误差分别为0.67和1.622 0,相对均方根误差为25.97%。该研究基于无人机获取的棉花冠层光谱反射率,从不同光谱预处理、波段筛选及建模方法建立的模型中筛选出最佳估算模型用于棉花全生育期LAI监测,研究结果可为棉花大田精准管理及变量施肥提供依据。

Monitoring of cotton leaf area index using machine learning

Leaf area index (LAI) is one of the most important indicators that characterize canopy structure and growth of crops. LAI changes can therefore greatly contribute to the variable rate fertilization of cotton. It is of great significance to monitor LAI quickly, accurately, and non-destructively, thereby guiding crop fertilization in modern agriculture. The traditional LAI monitoring relies mainly on manual sampling with high labor intensity and time-consuming. Furthermore, the lagging data cannot meet the needs of real-time monitoring. Most studies on crop LAI have also been made using remote sensing in recent years, such as hand-held spectrometers, unmanned aerial vehicles, and satellites. Nevertheless, the near-earth surface spectrum cannot be used to continuously and rapidly monitor at the spatial scale, due to the limited shooting range and the weight of the instrument. Satellite images are mostly used for the plant LAI monitoring at forest or large regional scale, particularly on the resolution of 10-60m. Alternatively, an unmanned aerial vehicle (UAV) has the potential to fast capture high resolution images repeatedly, suitable for accurate crop monitoring of small plots. Many efforts have been made to monitor the LAI of wheat, rice, corn and others using spectral images under UAVs. Since spectral technology can monitor timely and dynamically, and in macro mode, the resulting LAI spectral data really determines the vegetation index. As such, the hyperspectral reflectance of plant canopy can provide much richer information of vegetation characteristics, compared with vegetation index. However, a large amount of hyperspectral data under UAVs normally presents data redundancy and high multicollinearity. Reasonable spectral transformation can also be utilized to remove the background and noise of hyperspectral data. Correspondingly, machine learning has widely been applied to crop growth monitoring for deep information in data, particularly combined with remote sensing. Great ability of learning and prediction can be achieved using the partial least squares (PLS) model (an extension of multicollinearity model), support vector machine (SVM), and random forest (RF), in order to reduce the collinearity between variables in different ways. In this study, the UAV hyperspectral data was preprocessed using the first derivative (FDR), the second derivative (SDR), Savitzky-Golay(SG) smoothing, and multiple scatter correction (MSC) under the plot experiments of different varieties and nitrogen treatments. Sensitive bands were also selected using the Pearson correlation coefficient, successive projections algorithm (SPA), shuffled frog leaping algorithm (SFLA), and competitive adaptive reweighting (CARS). A cotton LAI monitoring model was finally constructed to calculate the reflectance of selected bands using the partial least square (PLSR), support vector regression (SVR), and random forest regression (RFR). The results showed that the canopy spectra of different LAI were significantly different from 760-1000 nm, where there was a significant correlation between the canopy spectrum and LAI. The sensitive response band of LAI in the cotton canopy was concentrated in the visible light (400-780 nm) and near-infrared (after 900 nm). The highest precision and stability were achieved in the RFR model under each pretreatment for LAI monitoring. Among them, the FDR-SFLA-RFR model performed the best, where the determination coefficient, root mean square error (RMSE), and relative RMSE for the modeling dataset were 0.74, 1.6483, and 26.39%, respectively. In the verification dataset, the determination coefficient, RMSE and relative RMSE were 0.67, 1.6220, and 25.97%, respectively. Consequently, the optimal estimation model can be rationally selected to represent the UAV spectral reflectance of the canopy using various pretreatments, band selecting, and modeling. The findings can provide the potential basis to accurately manage the variable fertilization in cotton fields.