基于特征降维和机器学习的覆膜冬小麦LAI遥感反演

为进一步提升无人机遥感快速监测覆膜条件下冬小麦叶面积指数(Leaf area index, LAI)的能力，以垄沟覆膜冬小麦为研究对象，利用无人机搭载五通道多光谱传感器获取2021—2022年冬小麦出苗期、越冬期、返青期、拔节期、抽穗期和灌浆期的遥感影像数据，使用监督分类剔除背景并计算50种可见光和近红外植被指数，采用主成分分析、相关系数法、决策树排序和遗传算法进行特征降维，结合偏最小二乘、岭回归、支持向量机、随机森林、梯度上升和人工神经网络6种机器学习算法建立不同输入特征变量下的覆膜冬小麦LAI反演模型，并进行精度评价。结果表明，剔除覆膜背景使冬小麦冠层反射率更接近真实值，提高反演精度。采用适宜的特征降维方法结合机器学习算法能够提高覆膜冬小麦LAI的反演精度和稳定性，对比特征降维前的反演精度，主成分分析和相关系数法无法优化反演效果，决策树排序只适用于基于树模型的随机森林和梯度上升算法，遗传算法优化效果明显，遗传算法-人工神经网络模型反演效果达到最优(决定系数为0.80,均方根误差为1.10,平均绝对值误差为0.69,偏差为1.25%)。研究结果可为无人机遥感监测覆膜冬小麦生长状况提供...

Remote Sensing Inversion of Leaf Area Index of Mulched Winter Wheat Based on Feature Downscaling and Machine Learning

To further improve the ability of UAV remote sensing to rapidly monitor the leaf area index (LAI) of winter wheat under mulching conditions, a UAV with a five-channel multispectral sensor was used to acquire remote sensing image data of winter wheat during the emergence, overwintering, rejuvenation, plucking, tasseling and filling stages from 2021 to 2022, using supervised classification to remove background and calculate 50 visible and near-infrared vegetation indices. The LAI inversion models of mulched winter wheat with different input feature variables were developed and evaluated in terms of accuracy by using six machine learning algorithms: partial least squares, ridge regression, support vector machine, random forest, extreme gradient boosting and artificial neural network. The results showed that removing the mulched background would make the reflectance of winter wheat canopy closer to the real value and improve the inversion accuracy. The inversion accuracy and stability of mulched winter wheat LAI can be improved by using a suitable feature reduction method combined with machine learning algorithm, and the inversion accuracy before feature reduction cannot be optimized by principal component analysis and correlation coefficient method, and the decision tree ranking was only applicable to random forest and extreme gradient boosting algorithm based on tree model, and the optimization effect of genetic algorithm was obvious, genetic algorithm-artificial neural network model inversion effect reached the optimal (R2 was 0.80, RMSE was 1.10, MAE was 0.69, and deviation was 1.25%). The research results can provide theoretical reference for UAV remote sensing to monitor the growth condition of mulched winter wheat.