特徵匹配於AVNIR-2影像與PALSAR影像 之影像套合

Abstract

中文摘要 遙感探測以非接觸的方式進行觀測，感測器主要分為被動式光學感測器和主動式雷達感測器，光學影像的成像原理為角度投影，並接收物體反射太陽的電磁波；雷達影像則使用距離投影成像，主動發射能量並接收地表反射的電磁波。兩種影像擁有不同的資訊，因此整合這些影像有利後續之應用，例如災害監控、地物分類、影像融合和變遷偵測等，然而影像整合的首要工作為影像套合。本研究以互訊息匹配 (Mutual Information Matching)、面特徵匹配(Patch-based Matching)及線特徵匹配(Edge-based Matching)三種方法進行異質影像套合，其中互訊息匹配利用移動視窗的熵 (Entropy)計算影像間的互訊息資訊量，互訊息資訊量最高者為相似性最高的區域；面特徵匹配是以分水嶺影像分割法進行面特徵萃取，並以成本函數(Cost Function)計算面特徵的相似度進行匹配；線特徵匹配是先以Canny萃取線特徵，再最小化兩組影像間邊緣線的距離，並以影像金字塔由解析度粗到細進行匹配。 本研究分別使用兩組AVNIR-2光學影像及PALSAR雷達影像，第一組資料位於日本東京灣，光學及雷達影像的影像等級分別為Level 1A和Level 1.1，第二組資料位於台灣北部，光學及雷達影像的影像等級分別為正射影像及Level 1.5，並就不同地表覆蓋物、地形起伏及影像品質衛星影像擷取測試區，分別使用上述三種匹配方法進行套合，最後以人工量測的檢核點作為精度評估的依據。實驗成果顯示，初步套合誤差約為25個像元的測試區一，互訊息匹配成果的均方根誤差約為4.27個像元，面特徵匹配為23.38個像元，線特徵匹配為3.41個像元；於山區地形的測試區二，初步套合誤差約為5.3個像元，互訊息匹配成果的均方根誤差值約為4.65個像元，面特徵匹配為4.51個像元，線特徵匹配為2.15個像元；於地表覆蓋物複雜的測試區三，初步套合誤差約為2.92個像元，互訊息匹配的成果的RMSE值約為2.38個像元，面特徵匹配為2.69個像元，線特徵匹配為2.62個像元；於河道區域的測試區四，初步套合誤差約為3.72個像元，互訊息匹配的成果的RMSE值約為3.76個像元，面特徵匹配為3.25個像元，線特徵匹配為3.21個像元，成果顯示三種方法均能有效提升異質影像之套全精度。

Remote sensing is a technology which acquired data that is not in contact. The sensors of remote sensing include optical active sensor and radar passive sensor. Optical image is an angular projection and receives the measure energy that is naturally available. Radar image is a distance projection and provides its own energy source for illumination. These sensors contain different information. The integration of these images may be beneficial to many applications such as disaster monitoring, classification, image fusion and change detection. Image registration is the most important part before data fusion. The proposed methods include mutual information matching, patch-based matching and edge-based matching. Mutual information (MI) matching uses the entropy of moving windows to determine the mutual information between images and find out the area with the highest mutual information as the most similar area. Patch-based matching uses watershed segmentation for region extraction. Then, it calculates the similarity between regions with a cost function. Edge-based matching uses Canny edge detector and minimizes the distance between edges on the pair of images. The image pyramid is also utilized to improve the result of registration from coarse-to-fine. This study contains two data sets. The test image are AVNIR-2 optical image and PALSAR radar image. Case one is an AVNIR-2 Level 1A image and a PALSAR Level 1.1 image which are located in Tokyo Bay. Case two is an AVNIR-2 orthoimage and a PALSAR Level 1.5 image which are located in the northern part of Taiwan. Three matching methods are applied to different images as well as different land covers. Then, we use the independent check points to assess their results. In test 1, the accuracy of initial registration is about 25 pixels, the root mean square error (RMSE) of MI, patch-based and edge-based matching are 4.27 pixels, 23.38 pixels and 3.41 pixels, respectively. Test 2 is a mountain area and the accuracy of initial registration is about 5.3 pixels. The RMSE of MI, patch-based and edge-based matching are 4.65 pixels, 4.51 pixels and 2.15 pixels, respectively. Test 3 covers complex surface and the accuracy of initial registration is about 2.92 pixels. The RMSE of MI, patch-based and edge-based matching are 2.38 pixels, 2.69 pixels and 2.62 pixels, respectively. Test 4 is a watercourse area and the accuracy of initial registration is about 3.72 pixels. The RMSE of MI, patch-based and edge-based matching are 3.76 pixels, 3.25 pixels and 3.21 pixels, respectively. The experiment indicates that these three methods can improve the results of the initial registration by a few control points manually. Keywords: satellite image, image registration, image matching, feature-based matching.