以深度學習萃取高解析度無人機正射影像之農地重劃區現況資訊

為改善農地重劃區圖、地不符之問題，乃透過現況測量將所測得的土地使用現況與數化完成的地籍圖進行套疊分析。目前執行現況測量時多以地面測量方法為之，然此種測量方式耗費大量人力、時間，且其測量成果通常較為局部，難以測得全域之現況資訊。相較於傳統地面測量，近年來航空攝影測量逐漸被應用於地籍測量相關領域中，其中無人機以其低成本、快速產製高解析度正射影像等特性，彌補傳統地面測量耗時、耗力的不足，惟該方法仍以人工方式獲取現況資訊。在影像分割領域，深度學習已取得較傳統影像處理更佳的表現，其透過卷積自動學習影像特徵，能夠在短時間內完成影像分割，具有自動化、高效率等優點。為此本研究採用ResU-net協助萃取高解析度正射影像中農地重劃區全域的現況資訊，並分析經過後處理的萃取成果應用於地籍測量相關作業之可行性。於模型訓練方面，除了使用高解析度正射影像外，另加入密匹配產製之數值高程資訊，且為了使模型有抗不同解析度的能力，嘗試使用不同高解析度的宜蘭正射影像作為模型的輸入資料，透過實驗比較網路的訓練成果。研究成果顯示在標籤資料涵蓋高程變化處時，加入高程資訊能些微提升模型的偵測能力。本研究使用宜蘭資料訓練的模型權重作為初始權重，再使用部分宜蘭與台中區域資料進行遷移學習，實驗成果證實使用遷移學習可提升效率，模型於宜蘭測試資料的精度F Score達0.73；台中測試資料的F Score達到0.86。另外，本研究計算地測現況點與深度學習萃取成果的平面位置較差，以《地籍測量實施規則》第73條規定進行分析，經統計得約80%資料符合規定，顯示應用深度學習搭配高解析度正射影像協助萃取農地重劃區現況資訊有其可行性。

關鍵字

地籍測量；現況測量；深度學習；影像分割；遷移學習

並列摘要

The results of detail survey are used for overlap analysis with digitized graphic cadastral maps to solve the problem of inconsistence between current land situation and cadastral maps. Currently, the detail data is mostly surveyed by theodolites and satellite positioning instruments; however, it is time-consuming and labor-intensive. Additionally, the surveying result is usually local data, and it is unable to obtain global terrain detail data. In recent years, aerial photogrammetry has been applied to cadastral survey, of which UAVs are increasingly being used as a low-cost, efficient system which can support in acquiring high-resolution data and bridge the gap between slow but accurate field surveys and the fast approach of conventional aerial surveys. However, these methods are based on manual measurement. In the field of image segmentation, deep learning has shown a higher accuracy than other image processing methods. By stacking multiple layers of neural networks, it can extract the image features in a short time, and carry out image segmentation. Furthermore, it is automatic and high-efficiency. Therefore, this study attempts to use ResU-net to assist in extracting global terrain detail information from high-resolution UAV orthoimages of farm land readjustment areas, and evaluate the feasibility of using the post-processing results in the detail survey. Except for the high-resolution orthoimages, the digital surface model (DSM) by dense matching was also used. In order to overcome the problem of different resolution, orthoimages and DSM with different ground sampling distance in Yilan were used as training data and analyzed the results of networks. The results showed that if the label data covered the elevation changes, adding DSM data by dense matching could promote the accuracy of detection. After the model had been trained by Yilan data, the study used part of data in Yilan, and part of data in Taichung as training data to fine-tune the model. The experimental results showed that using transfer learning could speed up the training time. The F Score in Yilan testing data was 0.73; Taichung testing data was 0.86. After post-processing of detected result from deep learning, the study calculated the planar location differences between the detail point positions from ground survey and those from deep learning, then performed the analysis based on the article 73 of “Rules for Implementation of Cadastral Surveys”, the result showed that about 80% data meet the accuracy requirement and it demonstrated the feasibility of using deep learning to assist in extracting global terrain detail information from high-resolution UAV orthoimages of farm land readjustment areas.