- 學位論文
一個對於高光譜與合成孔徑雷達遙測影像資料融合的模擬退火特徵齊一化波段選取方法
A Band Selection approach of Simulated Annealing Feature Uniformity for the Data Fusion of Hyperspectral and SAR Imageries
摘要
遙測（remote sensing）是利用衛星酬載或空載感測器，從空中探測地球表面之覆蓋物的技術。本論文中，將高光譜影像與合成孔徑雷達影像進行資料融合，並使用「模擬退火特徵齊一化波段選取（simulated annealing feature uniformity band selection, SAFU）」方法將高光譜影像進行波段選取。 高光譜影像因為波段眾多，因此如何在眾多波段中挑選較有意義的波段進而讓分類器得到較佳的分類正確率，成為重要課題之一。先前，曾有學者提出「模擬退火波段選取（simulated annealing band selection, SABS）」方法，而本篇論文提出的SAFU方法，是將高光譜影像各類別轉換成各自的相關係數矩陣，這些相關係數矩陣經過SAFU方法後，得到唯一的群聚特徵空間，此一群聚特徵空間能選取最有意義的波段，讓高光譜波段選取能更有效率的決定所選擇的波段，並由這些挑選出來的波段，得到較佳的分類辨識率。實驗結果顯示，利用資料融合因各個類別資料間的差異特性，可以大大提高分類正確率，本論文提出之SAFU方法能更快速的從高光譜影像中選取波段，並提高分類正確率。此外，本文並嘗試利用平行計算的方法來加速SAFU的執行效率，以期達到波段選取時能有更高的效率。
並列摘要
With the recent advances of state-of-the-art sensors, data initially developed in a few multispectral bands today can be now collected from several hundred hyperspectral and even thousands of ultraspectral bands. While images are continuously being acquired and archived, existing methodologies have proved inadequate for analyzing such large volumes of data. As a result, a vital demand exists for new concepts and methods to deal with high-dimensional datasets. In this paper,we fuse hyperspectral imaging and synthetic aperture radar imaging. We use Simulated annealing feature uniformity band selection (SAFU) from hyperspectral imaging feature extraction. Previously, scholars have put forward the “simulated annealing band selection” (SABS) . In this paper, we propose a novel feature extraction method, called simulated annealing feature uniformity (SAFU) band selection approach to improve the computational and the precise performances of the “clustered eigenspace / feature scale uniformity transformation” (CE/FSUT) of SABS method for clustering the CE features. It takes advantage of the special characteristics of SA to concentrate the CE feature sets of different classes into the most common feature subspaces. A distance measure based on SAFU is then applied to decompose the similarity for land cover classification purposes. Compared with the CE/FSUT method, the SAFU can group the CE feature sets of each different class in the same orders and can unify the feature scales of each different CE feature set at the same time. It can simultaneously group highly correlated bands of each different class into the same CE feature sets with higher effectiveness but lower computational loads. To demonstrate the advantages of the proposed method, we compared several different configurations categorized by the parameters of constructing SA annealing schedule. The performance of the propose method is evaluated by fusing MODIS/ASTER airborne simulator (MASTER) images and the Airborne Synthetic Aperture Radar (AIRSAR) images. Compared with conventional feature extraction techniques, SAFU evinced improved discriminatory properties, crucial to subsequent PBF classification. It made use of the potentially significant separability embedded in high-dimensional datasets to select a unique set of the most important feature bands. The experimental results showed that the proposed SAFU approach is effective and can be used as an alternative to the existing feature extraction method for the data fusion of hyperspectral data sets.