Learning Flexible Hyperspectral Features

Most researches in the hyperspectral field adjust feature extraction techniques as major dimensionality reduction tools. This is due to the high dimensionality problems. Feature extraction techniques are not limited to such purpose but extended to handle the changing spectral responses. To achieve that, these techniques transform the spectral response into a new domain where features are arranged according to specific criterion. Each technique extracts unique features that are totally different to that others extract. Besides, each technique has advantages and disadvantages regarding handling the highly mixed datasets and the small training sample size. Therefore, utilizing a technique than another may lead to significant information loss. To overcome this problem and derive flexible features, the proposed approach combines the resulting features of each extraction technique in one feature vector and employs a Support Vector Machine (SVM) to classify it. The feature vector consolidates the benefits of each individual technique and neutralizes their disadvantages. Minimum Noise Fraction (MNF), Principle Component Analysis (PCA) and Independent Component Analysis (ICA) have been used in the proposed approach. Experimental results show that the proposed approach overcomes the traditional feature extraction techniques.