基于无人机多光谱信息与纹理特征融合的小麦叶面积指数估测

Wheat Leaf Area Index Estimation Based on Fusion of UAV Multispectral Information and Texture Features

摘要:

叶面积指数(Leaf area index，LAI)是作物生长监测和产量预测的重要指标之一，为探究基于无人机多光谱技术的小麦LAI估测模型潜力，本文以小麦育种材料为研究对象，基于无人机平台获取小麦拔节期、孕穗期、抽穗期、开花期的多光谱图像，得到12种植被指数（Vegetation index，VI）及各波段的8种纹理特征（Texture features，TF）。然后，利用皮尔逊相关性分析方法筛选与LAI相关性较强的VI和TF，在优选2类特征基础上，利用递归特征消除法（Recursive feature elimination，RFE）筛选两者结合的综合特征（Comprehensive features，CF）。最后，基于3类特征，采用多元线性回归（Multiple linear regression，MLR）、支持向量回归（Support vector regression，SVR）、梯度提升回归（Gradient boosting regression，GBR）3种机器学习算法构建LAI估测模型，比较模型在各生育期的估测精度差异。结果表明：CF有效提高了小麦各生育期LAI估测精度；3种机器学习算法中，GBR更具稳定性，对3类特征均有较好的LAI拟合效果；以植被指数RVI、NDVI和纹理特征NIR_COR、R_MEA作为输入变量，结合GBR算法能够准确估测小麦LAI，所有时期训练集R2为0.91，RMSE为0.45，测试集R2为0.84，RMSE为0.67。本研究可为基于多光谱技术的小麦LAI估测提供应用参考。

Abstract:

Leaf area index (LAI) is one of the important indicators for crop growth monitoring and yield prediction. In order to explore the potential of wheat LAI estimation models-based on UAV multispectral technology, taking wheat breeding materials as the research object, the multispectral images were obtained-based on the UAV platform at jointing, booting, heading and flowering stages of wheat, and further calculated 12 vegetation indices (VI) and eight types of texture features (TF) in each band. Then, the Pearson correlation analysis method was employed to identify VI and TF which strongly correlated with LAI, and the recursive feature elimination method (RFE) was used to screen the comprehensive features (CF) on the bases of the preferred two types of features. Finally, based on the three types of features, three machine learning algorithms including multiple linear regression (MLR), support vector regression (SVR) and gradient boosting regression (GBR) were employed to establish LAI estimation models, and the estimation accuracy of the models was compared at different growth stages. The results showed that the CF effectively improved the accuracy of wheat LAI estimation models at each growth stages;among the three machine learning algorithms, GBR performed greater stability, and had better LAI fitting for the three types of features;specifically, the LAI estimation model-based on GBR, using vegetation indices RVI, NDVI, and texture features NIR_COR, R_MEA as input variables, performed best, with R2 of 0.91 and RMSE of 0.45 in the training set, R2 of 0.84 and RMSE of 0.67 in the testing set for all stages. The research result can provide an application reference for LAI estimation of wheat-based on multispectral technology.