空載光達(airborne LiDAR)是種主動式對地表的量測技術，利用直接地理對位轉換距離和角度，成為精準三維坐標的點雲。近年來，一種新興的光達稱為全波形(full-waveform)光達逐漸受重視，其可記錄完整光束回傳能量值，這些記錄能量值的集合稱為波形，使用者可藉由波形分析得到更完整的地表反射物理特性與細節變化，有助於地形三維重建及地物判識。 在處理全波形光達資料的過程，波形萃取與分析是必要的一環。本研究首先濾除背景雜訊，接著利用高斯函數配合二次微分法對波形進行擬合(fitting)，並對儀器與二次微分法為初始值之擬合成果進行精度評估。接著萃取光達波形參數，包括波寬(Width)、振幅(Amplitude)、背向散射截面(Backscatter cross-section)，再配合光達特徵包括正規化高程(normalized height)、強度(Intensity)與影像特徵綠指標(Greenness Index)，作為地物分類的特徵。地物分類使用簡易貝氏(Na?ve Bayes)與隨機森林(Random Forest)兩種分類器，並就分類成果評估精度，藉此探討全波形光達資料與一般空載光達資料於不同分類器的成效。 本研究結果顯示，使用二次微分法提供之初始值的擬合成功率較高，擬合誤差較小，且分類成果較佳。在地物分類的部分，全波形光達所提供的參數對於植物的分析的結果顯著。另外，相較於簡易貝氏分類器，以決策樹為基礎的隨機森林分類器較適合本研究的光達地物分類。 Airborne LiDAR is an active remote sensing system. It can transfer the distance and orientation to point cloud by direct georeferencing. A new generation of LiDAR system called Full-Waveform (FW) LiDAR which could receive the whole return signal (so-called waveform) in a ray has become popular recently. With FW LiDAR, users can obtain more information of objects by analyzing the waveform, and it is helpful for three dimensional reconstruction and land cover distinguishing. In the processing of FW LiDAR data, the waveform parameter extraction and analysis are the important steps. In this study, after eliminating the background noise, the Gaussian modeling function with second derivative method was used for waveform fitting. The result was compared to Gaussian fitting using the initial value provided by the instrument. Then the features extracted from the waveform, including width, amplitude, backscatter cross-section, traditional LiDAR features, normalized height and intensity, and greenness index from image were used for land cover classification. The classifiers used in this study were Na?ve Bayes and Random Forest and compared with each other. The experimental results indicate that using the second derivative method could provide higher fitting successful rate, smaller Root Mean Square Error (RMSE) and better classification result. The land cover classification results demonstrate that full-waveform features are helpful for distinguishing different vegetation targets and the decision-tree-based Random Forest classifier is more suitable for landcover classification of LiDAR data used in this study.
