應用共軛梯度演算法在條帶式合成孔徑雷達目標物特徵增強處理

Stripmap Mode SAR Target Feature Enhanced Using Modified CG Method

10.6574/JPRS.2006.11(2).3

蔣政諺(Cheng-Yen Chiang)；陳錕山(Kun-Shan Chen)；吳宗達(Tzong-Dar Wu)

合成孔徑雷達 ； 掃描式 ； 集點式 ； 特徵萃取 ； 目標物增強 ； SAR ； Stripmap mode ； Spotlight mode ； feature extraction ； target enhancement

航測及遙測學刊

11卷2期（2006 / 06 / 01）

143 - 159

繁體中文

目標物點狀散射點特徵在合成孔徑雷達目標物影像辨識上非常重要，因此找尋合適的萃取方法就十分關鍵，目標物點狀特徵的萃取原則在於維持主波瓣的強度，並降低旁波瓣的影響，並且能有效反應真實目標物散射點位置。本研究以增顯目標物點狀散射點為主要目標。 合成孔徑雷達影像依取像方式可分成兩類，焦點式和掃描式。本研究主要是由M. Çetin所提出針對焦點式合成孔徑雷達增顯地表反射係數的演算邏輯為基礎，其設計理念是以回波訊號方程式為基礎，建構成為演算法之轉換核心矩陣，再針對這個模型作最佳化的運算。本研究將上述方法擴張到掃描式合成孔徑雷達的資料上，並且由訊號接收數學式重新建構出一個新的演算法核心矩陣，方便未來研究上的應用。 本研究運用調控演算法(Regularization Algorithm)增顯散射點特徵，由於合成孔徑雷達的資料為複數資料，求解過程以複數共軛梯度為主要概念並配合準牛頓演算法(Quasi-Newton iteration)和柯列斯基分解(Cholesky decomposition)處理最佳化的問題配合影像切割的方式來完成。本研究使用RADARSAT高解析度掃描式合成孔徑雷達影像為測試資料，並配合地真調查的結果來佐證，除次之外也使用NASA/JPL AIRSAR全偏極資料為第二組例子。 研究中將此方法與傳統方法MV(Minimum variance)和MUSIC(Multiple Signal Classification)的效能作比較，展示散射點增強的結果並使用評鑑指標如目標物對背景雜訊比、主波瓣3dB寬和實際運算時間。結果顯示在經過規則化適當調整演算法參數之後，無論是在評鑑指標或是在統計機率密度分佈圖上，目標點旁波瓣的抑制和目標物與背景分離上皆有明顯的效果。

Target identification and recognition of SAR images require good feature selection and enhancement. Due to the coherent process, it is difficult to discriminate the SAR target feature properties simply using the shape, shadow, tone, color and texture, to name a few. The scattering center is one of the important properties for extracting the SAR feature. In this paper, we modified an algorithm based a conjugate gradient (CG) optimum method originally proposed for spotlight mode SAR images, in order to enhance the Stripmap mode SAR targets. First, we introduced to the background of scatter center enhanced for Spotlight mode from early papers. Second, we presented the SAR received signal model as the basic of CG method. Following, most importance is to introduce how to construct a kernel of CG method. To validate the effectiveness and efficiency of the modified method, a series of RADARSAT SAR images at fine mode were tested with ground truth available overpass the image acquisition. In addition to, we import the fully polarization SAR data from NASAIJPL AIRSAR acquired from south Taiwan when Sep. 27, 2000. We also compared the performance with MV (Minimum Variance) and MUSIC (Multiple Signal Classification) methods. Performance indices include target to clutter ratio, 3 dB mainlobe width and CPU time. From the logarithmic probability density distribution and column ordering plot of enhanced image and original image, it was demonstrated that modified method provides the best performance among the three methods besides CPU time.