由於部分遙測影像在某些特定地區存在光斑或雜訊，如：建築物附近，容易在地物分類上產生問題，而本研究的目為都市地區綠地覆蓋的偵測，即找出植生與非植生區域。然而植被在這些區域附近的像元多少會受到干擾，導致人工判釋上的困擾，因此我們將利用Sentinel-2影像當作參考，以人工選取的方式挑選屬於植生和非植生的像元，作為本次訓練和測試使用的地真資料（Ground Truth），並在經輻射校正過後目標影像中，選取研究範圍作為訓練樣本和測試樣本的來源。本實驗的研究區域為「竹南頭份都市計畫地區」，並選取該地區清晰無雲的影像，能降低訓練和測試時的誤差和提升準確性。實驗採用Deep ML、Spectral DeseNet和Spectral-Spatial DenseNet三種不同的深度學習方式進行模型訓練，並隨機選取樣本作為訓練和測試資料。最後將三種深度學習方式的分類結果和根據植生指數NDVI閾值所選取的分類結果作比較，檢驗模型分類能力。;In the particular situation, we might get noise from the high reflectance region e.g. Building area, and it exactly affect our classification result and accuracy in vegetation regions. Thus, the purpose of this research is to detect green space coverage in urban areas, that is, to identify Vegetation and non-Vegetation area. However, the pixels of vegetation near these areas that mentioned above will be disturbed to some extent, which will cause problems in manual interpretation. Therefore, we use the Sentinel-2 image as a reference, selecting the pixels as the ground truth data manually. The training samples and test samples were selected from the target image after radiation correction base on the ground truth. The research area of this experiment is the "Zhunan Toufen Urban Planning Area", the clear and cloudless images of this area are selected, which can reduce the error and improve the accuracy during training and testing. The experiment uses three different deep learning methods, Deep ML, Spectral DeseNet and Spectral-Spatial DenseNet for model training, and randomly selects samples as training and test data. Finally, the classification results of these three deep learning methods are compared with the classification results selected according to the NDVI and EVI thresholds of the vegetation index to test the classification ability of the model.
