在本研究中，提出了一種基於支援向量機（SVM）的新穎特徵線轉換進行資料降維，稱之為SVMFLE，並將其用於改進生成對抗網絡（GAN）在高光譜圖像（HSI）分類中的效果。GAN在許多應用中都成功展示了其強大的分類能力。然而，在GAN方法中，由於傳統的主成分分析（PCA）預處理步驟乃是基於線性轉換，無法有效地獲取非線性資訊，因此我們提出了SVMFLE此一非線性，且基於支援向量機的特徵線轉換來改善此一問題。我們提出的SVMFLE降維方法主要分成兩個階段執行。在第一個SVM階段，提取支援向量以獲得部分的組間分散度矩陣。然後，在第二個敏感性分析階段，使用最近鄰居（NN）分類器找到支援向量與原資料的組間分散度矩陣的最佳比例，然後最後獲得降維的特徵空間。由於所獲得的特徵空間比一般特徵空間更具代表性和區別性，因此可以提高GAN在HSI分類中的效果。我們透過三個標準數據集與具代表性的GAN方法進行比較以驗證SVMFLE結合GAN方法的有效性。根據實驗結果，我們所提出的SVMFLE結合GAN方法的分類效果優於原GAN，並且在Salinas，Indian Pines和Pavia University數據集中其準確度分別為96.26％，86.61％和89.22％。;In this study, a novel feature line embedding based on support vector machine (SVM), termed SVMFLE, is proposed for dimension reduction (DR) and applied to improve the performance of generative adversarial networks (GAN) in hyperspectral image (HSI) classification. The GAN has successfully shown its powerful discriminative capability in many applications. However, since the traditional linear-based principle component analysis (PCA) pre-processing step in GAN cannot effectively obtain the nonlinear information, the feature line embedding based on support vector machine method, SVMFLE was proposed to alleviate this problem. The proposed SVMFLE dimension reduction scheme was performed through two stages. In the first SVM stage, the support vectors were extracted for obtaining a part of between-class scatters. Then in the second sensitivity analysis stage, the nearest neighbor (NN) classifier was used to find a suitable combination of the between-class scatters, and then obtain the final dimension reduced feature space. Since the obtained feature space was much more representative and discriminative than the conventional ones, the performance of GAN in HSI classification could be improved. The effectiveness of the proposed SVMFLE with GAN scheme was measured by comparing with the state-of-the-art work on three benchmark datasets. According to the experimental results: the performance of the proposed SVMFLE with GAN is better than the state-of-the-art, and got the accuracy of 96.26%, 86.61%, and 89.22% in Salinas, Indian Pines, and Pavia University datasets, respectively.
