以影像控制區塊進行福衛二號衛星影像定位

Abstract

我國於2004年5月21日發射之福衛二號(FORMOSAT-2)，其全色態影像可達兩米之空間解析度，可達高解析度衛星影像（High Resolution Satellite Imagery, HRSI）之範疇，雖然解析度仍遜於低航高之航測影像，但衛星影像具有快速收集空間資訊和長時間觀測重複區域之優勢，使其具有執行傳統航測業務之潛力。 除非有精確的軌道參數提供感測器初始方位，否則使用衛星影像前，必須由地面控制點建構影像之物、像對應關係，研究中採用的幾何轉換模式包括多項式轉換(Polynomial Transformation)、平行透視投影模式(Parallel Perspective Model)、有理函數模式(Rational Function Model)和自率光束法平差(Bundle Adjustment with Self-Calibration)等，而本研究並引入完成空三解算之航照影像控制區塊做為控制資料，進行福衛二號影像之物、像對應關係解算，並利用具有重疊區域之立體像對，進行物空間交會決定點位之三維座標。 本文針對高雄楠梓地區之福衛二號Level 1A影像進行實驗測試，該實驗區地勢較為平坦(高程變化110公尺以內)，共採用兩張2005年1月(早期)拍攝之影像及一張2006年1月拍攝之影像，影像控制區塊由中華顧問工程司所產製，點位物空間三維精度約為0.2~0.5公尺，由成果得知感測器穩定度、控制點數量與幾何轉換模式皆對單像定位精度有所影響，而交會幾何亦會影響物空間交會精度(反應在高程精度)。實驗中自率光束法平差僅需19~30個控制點即可穩定求解，其單像物、像對應可達到次像元之精度，而物空間交會平面可達1.3~1.7公尺，高程3.2~6.7公尺；而二階、三階有理函數可得最佳之成果，但需60~70個地控點成果才趨於穩定，其單像定位精度皆可達次像元，物空間交會於平面精度可達1公尺，高程約3~4公尺，顯示福衛二號產製空間資訊產品之潛力。

Taiwan launched FORMOSAT-2 remote sensing satellite on May 21 2004. The 2-meter ground resolution on panchromatic mode can be classed as high resolution satellite imagery. Although the resolution of satellite imagery is worse than aerial photo which has lower flight height, but it has some advantages to execute the business of traditional photogrammetry, such as rapidly gathering the space information and repeatedly observing the same area. This research investigates the geometric transformation models, including self-calibration bundle adjustment, polynomial transformation, parallel perspective model and rational function model to construct the object-to-image relationship of FORMOSAT-2 imagery by employing control patches. The test area of the Level 1A FORMOSAT-2 imagery is in Nantsu, Kaohsiung and the control patches with positioning accuracy of 0.2~0.5m in the object space were provided by China Engineering Consultant Inc. The primary experiment results show that the stability of the sensor, the number of control patches, and the chosen geometric transformation model would mostly affect the accuracy of object-to-image transformation. With an adequate number (more than 100) of control points well distributed in the study area, the positioning accuracy in image space is reported to around 0.5~0.7 pixels when using 3rd order rational function model and bundle adjustment with self calibration. Furthermore, the 3D positioning accuracy through intersecting overlapping conjugate image points are evaluated to be around 1m for horizontal component while 3~4 m in height direction.