基于无人机多光谱实测数据的草地生物量反演模型比较

Comparative analysis of grassland biomass inversion models based on unmanned aerial vehicle multispectral data

应用无人机近地面遥感技术估算草地生物量是目前较热门的方法,但构建的反演模型类型、变量、算法差异较大。通过在内蒙古锡林郭勒获取的无人机多光谱影像提取出波段反射率、植被指数等变量,与实际获取的地面样方调查数据结合,构建并对比了8种最常用的参数与非参数方法构建的草地地上生物量预测模型,评估不同模型的精度和建模变量,以期能够优化得到最佳预测模型。研究结果表明8种模型中参数模型精度相对较低,非参数模型具有更高精度;参数模型中多变量的广义线性模型优于线性、对数和指数这3个参数模型;非参数模型中K近邻、支持向量机、极端梯度提升和随机森林4种模型的决定系数R^(2)都大于0.7,但随机森林模型相对更稳健,且模型变量数最少。建模变量中归一化植被指数和红波段反射率变量对生物量估算作用较大。综上,随机森林模型是较适用于内蒙古锡林郭勒地区草原无人机近地面遥感技术估算草地生物量的模型,然而在超参数调整、算法优化,以及植被多源变量筛选等方面还需要更深入的研究。

Using unmanned aerial vehicle(UAV)near-ground remote sensing technology to estimate grassland biomass is a popular method at present.However,the inversion model types,variables and algorithms are quite different.Modeling variables such as band reflectance and vegetation index were obtained by UAV multispectral images combined with the actual survey data of ground samples in Xilingol.The prediction models of grassland aboveground biomass constructed by eight most commonly used parametric and non-parametric methods were constructed and compared.The accuracy and modeling variables of different models were evaluated in order to optimize the best prediction model.The results showed that among the eight models,the accuracy of the parametric model was relatively low,and the nonparametric model had higher accuracy.The multivariable generalized linear model in the parametric model was better than the linear,logarithmic and exponential models.Among the nonparametric model,the model determination coefficients R^(2) of K nearest neighbor,Support vector machine,XGBoost and Random forests were all greater than 0.7 and the random forest model was relatively more robust and had the least number of model variables.Among the modeling variables,the normalized difference vegetation index and red band reflectance played important roles in biomass estimation.In summary,the random forests model is more suitable for UAV near-ground remote sensing technology to estimate grassland biomass in grasslands of Xilingol,Inner Mongolia.But the hyperparameter tuning and algorithm optimization,as well as vegetation multi-source variable selection and other aspects need more in-depth researches.