以深度學習自動偵測並協助萃取真正射影像中的建物邊界線

Applying Deep Learning to Automatically Detect and Extract Building Boundary Pixels from Aerial True Orthoimages

本研究聚焦於運用深度學習技術，提出一種直接從航空真正射影像中自動偵測建物邊界像素的方法並探討後續萃取處理的幾何精度。資料來源選擇台北市信義區作為訓練區域，並以建物類型、密度及分布與台北市信義區差異顯著的台東縣綠島鄉作為測試區域，以驗證模型在異質場景下之偵測能力。模型選用以Resnet34為骨架之U-Net（稱ResUnet34）為主要影像偵測框架，並以台北市信義區資料作為深度學習模型訓練資料，設計航空真正射影像RGB波段與經濾除雜訊的DHM（Digital Height Model）資料之多通道輸入組合，以結合光譜與高度資訊強化建物邊界偵測辨識能力。接著用台東縣綠島鄉的測試資料執行遷移學習，最後將模型輸出建物邊界像素區塊經骨架化處理萃取得到一像素寬的建物邊界，並評估其幾何精度。整體而言，本文所提出ResUnet34深度學習偵測建物邊界像素之方法搭配合適之建物標籤資料經後續細化處理，具備初步應用於建物邊界自動化萃取之可行性，但其建物邊界萃取之完整性仍待加強。

This study focuses on leveraging deep learning techniques to propose a method for directly detecting building boundary pixels from aerial true orthoimages, and further evaluates the geometric accuracy of the post-processing results. The training dataset is derived from the Xinyi District of Taipei City, while the test area is set in Ludao Township, Taitung County, a region with distinct differences in building types, density, and distribution from Xinyi District of Taipei City. For the ResUnet34 deep learning model was trained using data from Taipei City, incorporating both RGB bands from true orthoimages and filtered DHM (Digital Height Model) data as multi-channel inputs to integrate spectral and height information for improved boundary detection. Transfer learning experiments conducted on Ludao Township. The detected boundary pixels were psot-processed through skeletonization, and the geometric accuracy was assessed.
Overall, the proposed method with suitbale building boundary label data, followed by post-processing steps, demonstrates the preliminary feasibility of automated building boundary extraction. However, further improvements are needed to enhance the continuity and completeness of detected boundaries.