應用空載全波形光達資料於波形分析與地物分類

Abstract

空載光達是一種主動式對地進行測量的技術，可快速取得地表物之三維坐標，並提供高密度及高精度之三維坐標。而全波形光達技術中重要的發展，其可以記錄回波的連續波形，可藉由分析波形得到更多的地表反射物理特性與地表細節與變化，讓使用者擁有較豐富及完整的資訊，有助於地形重建及地物判識。 在處理全波形光達資料的過程中，波形萃取與分析是首先且必要的一環。本研究的第一階段，先進行資料整理與檔案格式讀取，接著利用對稱與不對稱兩種函數(高斯函數及韋伯函數)對波形進行擬合(fitting)，並進行原始資料與擬合成果兩者間的精度評估，討論與驗證不同擬合函數對於全波形光達訊號處理的適切性。經過精度評估後，萃取光達波形參數包括波寬(echo width)、振幅(amplitude)、背向散射參數(backscatter cross-section coefficient)，配合光達幾何參數包括高程(Z)、高程差(dZ)、回波數(echo number)、多重回波百分比(echo ratio)，提供為接下來分析應用的特徵。 取得波形的參數後，接著將這些萃取得的波形參數作為分類之用。所以本研究的第二階段，就萃取出的波形參數配合人工判識資料，使用支持式向量機(SVM)與隨機森林(Random Forest)兩種方法做地表物的特徵分類，並就分類成果進行精度評估，藉此探討使用全波形光達資料與一般空載光達進行分類與兩種分類法的優缺點。 本研究結果顯示，波形擬合分析的部分，使用不對稱函數的擬合成果較好，擬合誤差較小；但是在於波形峰值的位置與對稱函數的成果差異不大，因此高斯函數為一個簡單省時的擬合方式。而在地物分類的部分，全波形光達所提供的參數對於樹木的分析的結果顯著；相較於支持式向量機，隨機森林分類法的成果較為佳。

Airborne Lidar is an active remote sensing system. It can obtain the three dimensional coordinates effectively, and provide high density and high precision 3-D point cloud. Full-waveform (FWF) lidar is a new generation of airborne laser scanner which receives one dimensional continuous signal. It offers useful information about the structure of the target. Therefore, the analysis of received signal of FWF lidar and obtaining the implicit information is helpful for landcover classification. In the processing of full waveform Lidar data, the waveform parameter extraction and analysis are the important steps. The major objective of this study is to analyze the received waveform and extract its parameters. We select Gaussian distribution as a symmetric function and Weibull distribution as an asymmetric function in waveform decomposition. Then, we calculate several accuracy assessment indicators between raw waveform data and fitting function for quality assessment. We use echo width, amplitude, backscatter cross-section coefficient, elevation, elevation difference, echo number, and echo ratio as waveform parameter of classification. After waveform parameter extraction, we select Support Vector Machine (SVM) and Random Forests (RF) as classifier for landcover classification. This study employs echo width, amplitude, backscatter cross-section coefficients and other features for classification. Error matrix is used to compare the performance of the classifiers. The experimental results indicate that the accuracy of asymmetric function is slightly better than symmetric function. However, the extracted peak positions from the Gaussian and Weibull are very close. Moreover, Gaussian distribution is relatively simple and easy to implement in the waveform analysis. The result of landcover classification shows that waveform parameters are helpful for classification and Random Forests classifier is better than SVM in our study cases.