基于PROSAIL与机器学习的无人机多光谱遥感LAI估算模型

Estimating LAI Model from UAV Multispectral Remote Sensing Based on PROSAIL Model and Machine Learning

摘要:

叶面积指数（LAI）是表征作物冠层结构与生长状态的关键参数,对农田生态监测及精准农业管理意义重大。为提高复杂环境下LAI反演精度与稳定性,本文基于PROSAIL模型构建114210组冠层反射率与LAI配对样本,利用光谱转换后的模拟数据计算24种植被指数,经Pearson相关性分析筛选出8个高相关指数作为建模特征。选取KNN、RFR与XGBoost 3种机器学习方法构建LAI反演模型,以无人机实测多光谱反射率及对应LAI为输入,评估各模型在4个土壤含水率梯度（T1～T4）、3个关键生育期（拔节期、抽穗期、灌浆期）及全含水率、全生育期混合条件下精度与适用性。结果表明：XGBoost模型表现最优,非线性建模及泛化能力较强;RFR模型在中高含水率与冠层稳定期反演效果较好;KNN模型对土壤背景与特征扰动敏感,低含水率及早期生育期表现欠佳,混合条件下仍具一定稳定性。研究结果证实"物理模拟+特征筛选+机器学习建模"策略在多含水率、多时相农田环境中反演LAI的可行性,可为无人机多光谱遥感驱动的作物参数反演提供技术支撑。

Abstract:

Leaf area index (LAI) is a key biophysical parameter that reflects crop canopy structure and growth status, playing an essential role in ecosystem monitoring and precision agriculture. To improve the accuracy and robustness of LAI inversion under complex conditions, it generated 114210 paired samples of canopy reflectance and LAI using the PROSAIL radiative transfer model. From the simulated spectra, totally 24 vegetation indices were calculated, and eight with high correlation to LAI were selected through Pearson analysis as modeling features. Three machine learning algorithms, i.e., K-nearest neighbors (KNN), random forest regression (RFR), and extreme gradient boosting (XGBoost), were then applied to build LAI inversion models. These models were validated by using UAV-based multispectral reflectance and LAI observations across four soil moisture content (T1~T4), three growth stages (jointing, heading, grain filling), and integrated conditions combining all moisture levels and stages. Results showed that XGBoost achieved the best performance with strong nonlinear modeling and generalization ability. RFR performed well under medium-to-high moisture and stable canopy stages. KNN was more sensitive to soil background and spectral disturbances, with reduced accuracy under low moisture and early growth stages, but retained moderate stability under integrated conditions. Research findings demonstrated the feasibility of integrating physical simulation, feature selection, and machine learning for LAI inversion, providing methodological support for UAV-based multispectral remote sensing of crop parameters.