使用物空間半全域匹配法於多視角WorldView-2衛星影像

Abstract

伴隨著科技的發展，高解析度衛星影像不僅在空間解析度上可以達到數十公分等級，光譜解析度及時間解析度也持續進步中，本研究使用同一時期拍攝的多視角WorldView-2高解析度衛星影像，其全色態影像空間解析度為0.5公尺及，另有8個多光譜波段；影像匹配對於產製數值地表模型(Digital Surface Model, DSM)為重要的一環，其中，半全域匹配法(Semi-Global Matching, SGM)為匹配方法中最常使用的一個，其能夠在效率及精度上取得平衡，然而，傳統SGM最大的限制在於其僅能夠使用兩張核影像於像空間中進行匹配，為了解決這個問題，本研究提出OSGM (Object-based SGM)，OSGM與SGM最大的不同是OSGM在物空間中進行匹配，且可同時在物空間中計算多張影像的匹配成本值及進行匹配成本值加總。本研究使用OSGM法針對高解析度衛星的全色態影像及多光譜影像產製DSM。使用有理函數模型(Rational Function Model, RFM)以建立高解析度衛星影像之物像空間轉換關係，接下來建立以多影像為主的物空間影像匹配。首先針對不同的參數設置進行探討，探討最佳之參數設置原則；第二部份則比較SNCC及OSGM的DSM成果，以及不同匹配成本值之間的差異，研究中使用了三種匹配成本值分別是AD(Absolute Difference)、Census、NCC(Normalized Cross-Correlation)；第三部份探討有無使用影像金字塔之效率及成果；最後則針對多光譜影像的匹配成果，探討多光譜影像於物空間半全域匹配之效益分析。

Due to the development of satellite technology, high-resolution satellite images attain sub-meter spatial resolution and also improve its spectral and temporal resolutions. This study focus on the image matching for WorldView-2 multi-view satellite images. The WorldView-2 is capable to provide 0.5 m spatial resolution for panchromatic images and 2.0 m spatial resolution for 8 multi-spectral bands. As image matching is an important subject for producing digital surface model (DSM). semi-global matching (SGM) is widely adopted in dense matching due to its efficiency and accuracy. However, the traditional pair-wise SGM has a restriction that it could only perform matching with two epipolar images in image space. In order to overcome this issue, object-based SGM (OSGM) is proposed to perform image matching in object space. OSGM calculates matching cost for multi-images simultaneously and perform cost aggregation in object space directly. The research combines panchromatic and multi-spectral image of high-resolution satellite image (HRSI) and OSGM method to produce DSM. The WorldView-2 utilizes rational function model (RFM) for establishing the transformation between image and object space. Then, it performs object-based image matching for multi-view images. The experiment analyzes the impact of different parameters. Then, to compare the result of DSM using traditional object-based matching and OSGM. Besides, this research uses three kinds of matching costs (i.e. AD, Census and NCC). This study also compares the results of DSM with and without image pyramid strategy. Finally, different multispectral bands are adopted in OSGM.