基于随机森林模型的林地叶面积指数遥感估算

Estimation of Forest Leaf Area Index Based on Random Forest Model and Remote Sensing Data

林地叶面积指数(Leaf area index,LAI)的准确估测是精准林业的重要体现。为了快速、准确、无损监测林地LAI,利用LAI-2200型植物冠层分析仪获取福建省西部森林样地的LAI数据,结合同期Pleiades卫星影像计算12种遥感植被指数,分析了各样地实测LAI数据和相应植被指数的相关性,进而使用随机森林(RF)算法构建了林地LAI估算模型,以支持向量回归(SVR)模型和反向传播神经网络(BP)模型作为参比模型,以决定系数(R^2)、均方根误差(RMSE)、平均相对误差(MAE)和相对分析误差(RPD)为指标评价并比较了模型预测精度。结果表明:全样本数据中,各植被指数与对应LAI值均呈极显著相关(P<0.01),且相关系数都大于0.4;RF模型在3次不同样本组中的预测精度均高于同期的SVR模型和BP模型;3个样本组中RF模型的LAI估测值与实测值的R^2分别为0.688、0.796和0.707,RPD分别为1.653、1.984和1.731,均高于同期SVR模型和BP模型,对应的RMSE分别为0.509、0.658和0.696,MAE分别为0.417、0.414和0.466,均低于同期其他2种模型。

Accurate estimation of forest leaf area index （LAI） , which is defined as half the total area of green leaves per unit ground surface area, is the important embodiment of precision forestry. In order to monitor forest LAI faster, more accurate and non-destructively, LAI- 2200 plant canopy analyzer was used to acquire LAI data from the forest plots in western Fujian. Totally 12 kinds of vegetation index based on the Pleiades satellite images in the same period were calculated and the correlation between measured LAI and the vegetation index was analyzed. The purpose was to construct LAI estimation model specifically by using random forest algorithm （RF）. Additionally for each sample group, the models based on support vector regression model （SVR） and back-propagation neural network model （BP） were employed as comparison models. The estimation accuracy of the three models for each sample group was compared based on determination coefficients （R2）, root mean square errors （RMSE）, mean relative errors （MAE） and relative percent deviation （RPD）. The results indicated that the vegetation indices and LAI values were significantly correlated （P 〈 0.01 ） , and the correlation coefficients were greater than 0.4 for all sample data. The forecast accuracy of RF model in three different sample groups was higher than those of the SVR and BP models in the same period. R^2 of LAI estimated and measured values in the three sample groups based on RF model were 0. 688, 0. 796 and 0. 707, respectively. RPD were 1. 653, 1. 984 and 1.731, respectively. These data were all higher than those of SVR model and BP model, and RF model showed a higher accuracy than the other two models （RMSE of RF model were 0. 509, 0. 658 and 0. 696, respectively; MAE were 0. 417, 0. 414 and 0. 466, respectively）. These results would be helpful for improving the forest LAI remote sensing estimation accuracy.