| 摘要: | 近岸水底地形模型對於研究海洋生物的空間分佈、監測海底形態的變
化及產生精確之正射校正影像是很重要的資訊。可獲取水下地形的方法中，聲納、空載測深光達精度高但須至現地量測。而衛星影像推算水底地形，適合用於無法到達和有爭議的區域。其中，衛星影像的光譜資訊可以幫助推估反演地形模型，但需要精確的訓練數據。另一方面，數位攝影測量可使用高解析度衛星立體對直接測量精確的水下地形，不需要訓練資料。因此，本研究利用高解析度衛星立體對產製水下地形。
本研究方法的概念為先利用初始地形模型產製的兩張正射影像進行匹配。若地形模型正確無誤，則兩張正射影像應極為相似。然而，若匹配所得之共軛點具有視差，則代表地形模型具有誤差。過去文獻常使用外方位參數推估共軛點視差對應之高程修正量，但現在常見的高解析度衛星資料僅提供有理函數模型，無法直接得知衛星拍攝的位置及姿態。因此本研究使用立體對的交會角、等分線仰角和不對稱角，配合匹配取得之視差估算高程修正量。
提出的方法包括四個主要步驟：（1）前置處理，（2）計算高程修正量，（3）地形重建，（4）水下地形折射校正。首先，使用初始水下地形模型產生待匹配之正射影像，為了增加影像匹配的效能，我們只匹配從主正射影像中提取的特徵點。再計算共軛點的視差，估計地形模型上的高程修正量，此過程配合影像金字塔迭代。最後，由於光折射的性質，需要進行折射校正以產生最終的水底地形模型。
本研究之測試區位在南海的東沙環礁。使用空載測深光達產生的地形模型進行精度檢核後，有以下幾點發現：（1）使用融合影像之匹配點精度在淺區域精度高，可達0.52 公尺、（2）對於地形模型，由於綠波段影像於深區域之成功匹配點較多，模型整體精度優於融合影像之成果，可達1.15公尺、（3）水底地形特徵不明顯，小的匹配視窗（小於31×31）容易造成錯誤匹配、（4）綠波段之水體穿透性佳，不同匹配門檻（0.6, 0.7, 0.8）之設定無明顯的成果差異。
;Coastal digital elevation model (DEM) is important to map the spatial distribution of marine organisms, monitor changes of seafloor morphology, and produce accurate orthorectified images. There are different approaches for
bathymetry mapping. For example, sonar and airborne bathymetric Lidar have high accuracy, but both face difficulties on monitoring inaccessible and controversial area. On the other hand, satellite imagery does not have this limitation. The spectral information in satellite imagery can be helpful for retrieving coastal DEM. However, this approach requires a good quality of
training data. Therefore, digital photogrammetry approaches are more preferable as they can measure accurate bathymetry without the training data requirement.
This research first uses an initial DEM to generate two orthorectified images for image matching. If the DEM is accurate, these two orthorectified images should be very similar. However, if parallaxes happen between the orthorectified images, we assume they are caused by the incorrectness of the DEM. As the traditional approaches often use the exterior orientation parameters (EOPs) of images to estimate elevation corrections, EOPs may not be available for every satellite images nowadays. Hence, this research estimates the elevation corrections from parallaxes by using the convergence angle, bisector angle, and asymmetry angle of the stereo-pair.
In general, the proposed method comprises four main steps: (1)pre-processing, (2) elevation correction, (3) DEM reconstruction, and (4) refraction correction. First of all, an initial DEM is applied to produce orthorectified images. In order to increase the performance for image matching, we only match the features extracted from the master image. After calculating parallaxes, we can estimate the elevation corrections and iterate the process using image pyramids. Finally, because of the refraction effect, the refraction correction is necessary to produce the final bathymetry DEM.
We have examined the proposed solution on the Dongsha Atoll in the South China Sea. By comparing with a DEM derived from Lidar, we have the following observations: (1) For the accuracy of matched points, pan-sharpened image has better performance on the shallow water region, which is about 0.52 meters. (2) For the accuracy of DEM, green bend images can achieve more match points on the deeper water region, which result in a more accurate (i.e., 1.15 meters) DEM. (3) Since underwater features are less obvious, small target window size (i.e., less than 31 × 31) would result in wrong matches. (4) In terms of the correlation coefficient threshold, as the Green band has good water penetration performance, there were no significant difference when using different thresholds (i.e., 0.6, 0.7, 0.8). |
