應用全卷積網路於遙測影像偵測都市地區之建築物與道路

都市地區的土地利用圖是提供整體都市規劃與管理的重要空間資訊。儘管已經開發許多從遙測影像中來萃取都市土地利用的分類方法，但是這些傳統方法受限於不同影像的解析度，以及分類的準確性與效率，不足以滿足現實世界中的應用需求。近年來，深度學習的分類方法興起，已經在影像辨識中達到相當高的性能水準。因此，本研究使用基於深度學習之全卷積網路（Fully Convolutional Network, FCN），從無人飛行載具（Unmanned Aerial Vehicles, UAV）和衛星所取得不同解析度之RGB影像中，萃取都市地區（高雄市鳳山區）之建築物及道路。基於FCN的架構重新設計網路層數與結構，並進行各種超參數的調整與測試，將其學習到的特徵轉換至分割任務。本研究所建構FCN模型之學習誤差小於0.073，且無產生過擬合現象。實驗結果顯示，該模型對於UAV影像或衛星影像之整體準確度（overall accuracy, OA）均高於97%以上，可以有效提供都市地區地圖更新之參考。

關鍵字

深度學習； UAV影像；衛星影像；建築物；道路

並列摘要

The Land-use maps of urban areas present important spatial information that informs urban planning and management. Although many classification methods for remote sensing images have been developed to derive land-use information in urban areas, traditional methods are limited by image resolution, and in many cases, the accuracy and efficiency of legacy classification processes are not sufficient for real-world requirements. Recently, deep learning classification methods have emerged and improved, attaining a high level of performance in image recognition. This study employs those advances, adopting a fully convolutional network (FCN) based on deep learning to extract buildings and roads in the selected area (Fengshan District, Kaohsiung). This application uses RGB images of different resolutions obtained from Unmanned Aerial Vehicles (UAVs) and satellites. The FCN-based architecture revises the number and structure of network layers and performs various hyperparameter tuning and testing to translate learned features into segmentation tasks. The learning error of the FCN model constructed in this study is less than 0.073, and there is no overfitting phenomenon. The experimental results show that the overall accuracy (OA) of the model for UAV images and satellite images is higher than 97%, which can effectively provide a reference for updating maps in urban areas.

並列關鍵字

deep learning ； UAV image ； satellite image ； buildings ； roads

參考文獻

Chen, J., Y. Zhou, A. Zipf, and H. Fan. 2019. Deep learning from multiple crowds: A case study of humanitarian mapping. IEEE Transactions on Geoscience and Remote Sensing 57 (3): 1713-22, doi: 10.1109/TGRS. 2018.2868748

Feng, Q., J. Liu, and J. Gong. 2015. UAV remote sensing for urban vegetation mapping using random forest and texture analysis. Remote Sensing 7 (1): 1074-94, doi: 10.3390/rs70101074

Gibril, M. B. A., H. Z. M. Shafri, and A. Hamedianfar. 2017. New semi-automated mapping of asbestos cement roofs using rule-based object-based image analysis and Taguchi optimization technique from WorldView-2 images. International Journal of Remote Sensing 38 (2): 467-91, doi: 10.1080/01431161.2016.1266109

Gómez-Chova, L., D. Tuia, G. Moser, and G. Camps-Valls. 2015. Multimodal classification of remote sensing images: A review and future directions. Proceedings of the IEEE 103 (9): 1560-84, doi: 10.1109/JPROC.2015.2449668

Guo, L., N. Chehata, C. Mallet, and S. Boukir. 2011. Relevance of airborne lidar and multispectral image data for urban scene classification using Random Forests. ISPRS Journal of Photogrammetry and Remote Sensing 66 (1): 56-66, doi: 10.1016/j.isprsjprs.2010.08.007

主題瀏覽