特徵分析於光譜與光達特徵之物件式分類

Abstract

遙測資料有效提供地表覆蓋資訊，並應用於土地管理與監測。遙測影像資料具二維之覆蓋物光譜資訊，而光達點雲提供了良好的三維空間資訊，整合兩項互補遙測資料得提供覆蓋物良好判釋依據；此外物件式分類，基於影像物件進行分類，較像元式分類可提供更完整的地物特性。 整合多種遙測資料可獲得多元的分類特徵，如整合多光譜與光達特徵，或整合高光譜影像與光達特徵；此外多元分類策略提供更有彈性的分類架構，可提升特徵資訊的效益。本研究整合光譜與光達特徵，結合支持向量機與決策樹之物件式多元分類，並分析各項光達特徵，及不同光譜解析度之光譜特徵對分類效益的影響。 實驗資料包含ITRES CASI 1500高光譜影像、WorldView-2多光譜影像、及Optech ALTM Pegasus光達點雲。光譜包含MNF、NDVI及紋理影像；光達特徵包含nDSM、強度值及回波率。本研究的分類程序如下：設定九項分類目標；針對目標類別圈選訓練區；根據訓練區位置進行特徵分析，確定各目標之特性與適用的特徵；根據目標特性，建立結合支持向量機之分類決策樹，其中影像物件採重新分割與繼承式分類策略產生。並於最後針對分類成果進行比較分析。本研究分類實驗採用的特徵組合，主要分為(1)僅高光譜特徵、(2)高光譜整合單一光達特徵及(3)多光譜與高光譜各別整合光達特徵。 實驗成果提供分類成果展示，並針對各項特徵組合之分類精度進行比較與分析。相較於多光譜特徵，高光譜特徵提升10%~15%的整體精度；整合單一光達特徵與高光譜特徵後可提升4%~10%的整體精度，並且在整合所有光達特徵與高光譜特徵後，可達約15%之精度提升。單一光達特徵於精度指標的表現，強度值對植被等具反射差異的位置有較明顯提升，最高約3%提升量；回波率則對結構特性較明顯的目標物提供較多改善，最高約5%提升量；nDSM則在高度差異明確的目標表現較佳，最高約5%提升量。

Remote sensing technologies efficiently obtain information such as images and lidar point clouds of land covers to assist land use management. Images provide 2-D spectral information; while lidar point clouds present 3-D spatial information of land covers. The two complementary data can be beneficial for cover identification if they are appropreately integrated. Moreover, compared with pixel-based classfication, the object-based classification features more complete shape for land cover identification. Integrating multisource data of remote sensing, for example, integration of multispectral/hyperspectral image and lidar, enables different types of features and wider applications. Furthermore, the hybrid approach classification is considered more flexible than traditional ones. Given the above advantages, this study aims to combine the spectral and lidar features to creat hybrid approach and object-based classification so that to illuminate merits of each under different features and spectral resolutions. The examined data included hyperspectral imagery from ITRES CASI 1500, multispectral imagery from WorldView-2, and lidar data from Optech ALTM Pegasus. Spectral features were MNF, NDVI, and texture image; Lidar features included intensity, echo ratio, and nDSM. The proposed scheme incorporated four main steps: (1) defining 9 land covers in the studied area, (2) training area selection, (3) feature assessment, and (4) establishing decision tree with Support Vector Machine (SVM) integrated. Moreover, the re-segmentation and inheriting classification were used to produce image objects. The combinations of features were: (1) classification using hyperspectral features only, (2) classification using hyperspectral and each lidar features, and (3) classification using lidar and different spectral features. The classification results were revealed in both quantitative and qualitative forms. The experiment results suggested 10% to 15% improvement of overall accuracy for hyperspectral features; while for hyperspectral features with different lidar features, it was 4% to 10% increase. Moreover, integrating hyperspectral features and all lidar features resulted in 15% growth of overall accuracy. To compare each lidar feature, the improvement for intensity, echo ratio, and nDSM were 3%, 5%, and 5%, respectively. In general, intensity improves the separation of vegetation and other land covers; echo ratio identifies land covers with structural characteristics; nDSM separates land covers by different height characteristics.