監督式機器學習於土地覆蓋分類效益之研究

無人飛行載具（Unmanned Aerial Vehicles, UAV）之遙測影像相較於衛星影像有快速、機動取得地表資訊之能力，並具有低成本、高空間與時間之解析度，以及影像資料較不受雲霧干擾之特性，已廣泛地運用在小區域之監測與調查作業。本研究運用UAV高效率的遙測取像方式，並結合支持向量機（Support vector machine, SVM）、最大概似法（Maximum likelihood, ML）及隨機森林（Random forest, RF）三種監督式機器學習方法實施地表特徵樣本訓練及測試，再評估五種土地覆蓋（樹木、草地、裸露地、建築物及道路）之分類效益。旨在比較和找到最合適的分類器，以有效率地用於UAV影像之土地覆蓋分類。在鄉村地區研究結果顯示SVM的分類準確率為88%、曲線下面積（Area under the curve, AUC）為0.88、Kappa值為0.83及Gain為96.8%（前50%測試集），其綜合評估的分類效益最佳。另外，選擇地物較複雜的都市地區進行測試，SVM的分類準確率為85.4%，也是三種分類器中最佳的，尤其對於道路能正確地預測（分類）。本研究所使用之機器學習是基於RGB做出預測，無論是在鄉村或都市地區的土地覆蓋分類均有良好的成果，且三種監督式機器學習（分類器）準確率都大於78.6%以上。整體而言，三種分類器能清楚區分各種土地特徵的差異，並分析人為（building、road）與自然（tree、grassland、land）的不同光譜組成與特性，且正確的執行土地覆蓋分類。

關鍵字

土地覆蓋；分類；無人飛行載具；監督式機器學習

並列摘要

Compared with that realized through satellites, remote sensing images conducted using unmanned aerial vehicles (UAV) can yield land surface information more promptly and flexibly. Moreover, this sensing involves a low cost and has a high spatial and temporal resolution. In addition, the obtained image data involve less interference pertaining to clouds and fog. UAVs have been widely used in small area monitoring and investigation operations. In this study, the high-efficiency remote sensing image method based on UAVs is adopted, and three supervised machine learning methods, namely, support vector machine (SVM), maximum likelihood (ML), and random forest (RF), are combined to implement training and testing of the land surface feature samples. Subsequently, the classification benefits of five types of land cover (tree, grassland, land, building, and road) are evaluated to identify the most suitable classifier to be used for efficient land classification for the images obtained using the UAV. For the SVM in rural areas, the classification accuracy, an area under the curve (AUC), Kappa coefficient, and Gain are 88%, 0.88, 0.83, and 96.8% (first 50% of the test set), respectively. This classifier achieves the highest classification benefit. Next, a city area with more complex features is selected for testing. The SVM classification accuracy is 85.4%, which is the maximum among the three classifiers. In particular, the SVM classifier can accurately predict (classify) roads. The machine learning approach performs predictions based on RGB. Satisfactory land classification results are obtained both in rural and urban areas. The accuracy of all three supervised machine learning classifiers is greater than 78.6%. In general, all the classifiers can clearly distinguish the land features, analyze the different spectral compositions and characteristics of artificial (building and road) and natural (tree, grassland, and land), and accurately perform land cover classification.