運用空載高光譜及光達資料建立森林覆蓋分類判釋模型

本研究以林業試驗所六龜試驗林為範圍，結合空載高光譜及光達資料，建立地物／林型與特定樹種的光譜及三維結構特徵，並應用統計及機器學習演算法建立自動化分類判釋模型。模型候選變量包含高光譜代表波段、光譜衍生植生指標以及光達衍生冠層結構指標等三大類共19個變量，並透過相關分析排除冗餘變量後，篩選8個變量進入模型。變量重要性評估顯示，冠層高度是判斷地物／林型類別的重要結構特徵。雖然特定樹種分類模型入選的預測變量包含更多的結構指標，但重要性小於植生指標。在地物／林型分類中，支持向量機與隨機森林模型的整體分類精度可達75％，兩者整體精度差異為0.24％，且分類類別間的混淆情況相似。特定樹種分類以隨機森林模型整體精度最高（86.79％），其次為支持向量機（85％）。最大概似模型在地物／林型分類及特定樹種分類的表現均不佳。非參數統計的機器學習模型，其不假設資料統計分佈的特性更適合不同的感測器資料或輔助變量進行分類的目的，在特徵空間複雜的情況下也能得到穩健的分類結果。整體而言，整合高光譜訊息及光達衍生結構變量的機器學習模型能有效的分類更細緻的森林覆蓋類型。

關鍵字

高光譜影像；光達；機器學習模型；森林覆蓋分類

並列摘要

In this study, airborne hyperspectral imagery and LiDAR data were combined to establish spectral and 3-dimensional structural characteristics of land cover and forest types in the Liukuei Experimental Forest (LEF). Statistical and machine learning algorithms were used to develop automated classification models. In total, 19 variables were prepared as model candidate variables, which were divided into three major categories, including representative hyperspectral bands, vegetation indices calculated using hyperspectral data, and canopy structural indices derived from LiDAR data. Redundant variables were excluded by a correlation analysis, and the models were determined using 8 variables. Assessment of the importance of the variables showed that canopy height was an important structural feature for interpreting the land cover/forest types. Although more structural indices were included among the predictor variables selected by the specific tree-species classification model, they were less important than the vegetative indices. For the land cover/forest type classification, the difference between the overall accuracy of support vector machine (SVM) and random forest (RF) model was 0.24%. Both models yielded an overall accuracy of 75% with similar levels of confusion between classification categories. For specific tree-species classification, the overall accuracy of RF was the highest (86.79%), followed by SVM (85%). The maximum likelihood classification (MLC) had relatively poor performance in both land cover/forest type classification and specific tree-species classification. These non-parametric machine learning models, which do not rely on data following particular statistical distribution, are more suitable for classification purposes when using data from different sensors or auxiliary variables. Their classification accuracy was more robust than traditional classification techniques such as MLC, especially when the feature space is complex. Overall, machine learning algorithms that integrate hyperspectral information and LiDAR-derived structural variables can effectively distinguish more-detailed forest cover types.

並列關鍵字

hyperspectral imagery ； LiDAR ； machine learning model ； forest cover classification

主題瀏覽