基于无人机多光谱遥感的马尾松林叶面积指数估测

Leaf Area Index Estimation of Masson Pine (Pinus massoniana) Forests Based on Multispectral Remote Sensing of UAV

摘要:

快速、准确、无损估测马尾松林叶面积指数对精准林业管理具有重要意义。以小型低空无人机为平台，搭载RedEdge多光谱传感器，获取福建省西部马尾松林多光谱影像，运用重采样的方式获取并计算不同空间分辨率（0.08、0.1、0.2、0.5、1、2、5m）下的植被指数，结合地面实测LAI数据，分析其与植被指数的相关性，进而采用线性模型（LR）、多元逐步回归模型（MSR）、随机森林模型（RF）、支持向量机模型（SVM）和人工神经网络模型（BP）构建不同空间分辨率下的马尾松林LAI估测模型，以决定系数（R2）、均方根误差（RMSE）、相对分析误差（RPD）和总体精度（TA）来评价估测模型精度，从而确定最佳空间分辨率和最佳模型。结果表明，不同空间分辨率下LAI与植被指数均呈极显著相关（p<0.01）；多变量模型（MSR、RF、SVM、BP）的调整R2平均值高于LR模型；随着空间分辨率的增加，不同模型的R2整体上呈先增大后减小的趋势；当空间分辨率为0.5m时，利用植被指数建立的RF模型为马尾松林LAI的最佳估测模型，RF模型的调整R2为0.766，模型估测的R2、RMSE、RPD和TA分别为0.554、0.421、1.523和81.95%。本研究可为无人机多光谱遥感反演森林LAI表型参数的空间分辨率和模型选择提供理论参考。

Abstract:

Fast, accurate and non-destructive estimation of the leaf area index (LAI) of Masson pine forest is of great significance for precise forestry management. In order to estimate LAI of Masson pine forest, the small low-altitude unmanned aerial vehicle (UAV) platform with the American MicaSense RedEdge multi-spectral sensor was used to obtain the multi-spectral image in western Fujian. Eight different kinds of vegetation indices, green normalized vegetation index (GNDVI), green ratio vegetation index (GRVI), modified soil adjusted vegetation index (MSAVI), normalized difference vegetation index (NDVI), renormalized difference vegetation index (RDVI), ratio vegetation index (RVI), structure insensitive pigment index (SIPI) and visible atmospherically resistant index (VARI) were calculated from imagines with seven spatial resolutions (0.08m, 0.1m, 0.2m, 0.5m, 1m, 2m and 5m). The correlation between groundmeasured LAI and different vegetation indices from different spatial resolutions imagines were analyzed. Five models, linear regression (LR), multiple stepwise (MSR), random forest (RF), support vector machine (SVM) and artificial neural network (BP) were used to construct LAI estimation model, and coefficients of determination (R2), root mean square errors (RMSE), residual predictive deviation (RPD) and total accuracy (TA) were used to determine the optimal spatial resolution and optimal model for computing Masson pine forest LAI. The results showed that LAI values and vegetation indices from different spatial resolutions imagines were significantly correlated (p<0.01). The adjusted R2 average values of the multivariate models (MSR, RF, SVM, BP) were higher than that of the LR model. The R2 of different models was generally increased first and then decreased with the increase of spatial resolution. RF model was the optimal model for Masson pine forest when the spatial resolution was 0.5m. The highest accuracy for RF model with R2 of 0.766 for calibration, and with R2 of 0.554, RMSE of 0.421, RPD of 1.523, and TA of 81.95% for validation. The research result can provide a theoretical reference for the spatial resolution and model selection in the inversion of forest LAI phenotypic parameters by UAV multi-spectral remote sensing image.