自我學習提升效能之機器學習演算法逐漸普及於遙測影像資料處理及分析,機器學習的優勢在於不需預先了解全部資料的特性,且資料分佈不必為常態分布,較能描述遙測影像資料之實際分布。具高維度特性的資料,如高光譜影像,其重要資訊主要集中在較低維度的子空間中;此外,同類像元通常會近似分布於同一個低維度子空間中。如何降低資料維度並進行分類,成為高光譜影像分析的主要研究議題。近十年來,大量研究致力於以稀疏表示方法進行高光譜影像分類。稀疏表示在訊號重建方面有著良好的效能,其可處理具稀疏特性的資料,因此很適合運用於高光譜影像之處理及分析。 本研究將以聯合稀疏表示為基礎,提出一個適合高光譜影像分類之方法。在進行研究時,主要可分為三部分進行探討:一、字典之建構,透過機器學習演算法,利用已知訊號之訓練樣本獲得字典。二、稀疏係數最佳解之求解,以正交匹配追逐法等進行稀疏係數之求解。三、建立聯合稀疏表示之分類模型,同時整合空間及光譜資訊至稀疏表示演算法的目標函數中,建立以聯合稀疏表示為基礎的影像分類方法,期能提升高光譜影像分類之效能及準確性。
Machine learning algorithms using self-learning to improve performance have become increasingly popular in the processing and analysis of remote sensing image data. The advantage of machine learning is that it is not necessary to know the prior characteristics of data in advance, and the data distribution does not have to be normally distributed, so it is more able to describe the actual distribution of remote sensing image data. Most important information of high-dimensional data (e.g. hyperspectral images) is mainly clustered in low-dimensional subspace. Moreover, pixels belonging to the same class are usually distributed in the same low-dimensional subspace. Therefore, how to reduce the dimensionality for classification has become the major issue for hyperspectral image analysis. Considerable researches have been dedicated to hyperspectral image classification via sparse representation methods over the past decade. Sparse representation has shown good performance in signal reconstruction and can be used to process data with sparse properties, so it is quite suitable for hyperspectral images analysis. On the basis of sparse representation method, the paper consists of three main parts for discussion. First, the method for dictionary construction will be introduced. With machine learning algorithm, dictionaries can be obtained by training samples of provided spectral signal. Second, the solutions for sparse coefficients optimization, such as orthogonal matching pursuit is tested for experiment analysis. Third, the model of joint sparse representation for hyperspectral image classification will be put forward. In the proposed model, the spectral and spatial information are integrated into the joint sparse representation simultaneously to improve the efficiency and accuracy of the hyperspectral image classification.