Spectral–spatial multi-feature-based deep learning for hyperspectral remote sensing image classification

Hyperspectral remote sensing has a strong ability in information expression, so it provides better support for classification. The methods proposed to deal the hyperspectral data classification problems were build one by one. However, most of them committed to spectral feature extraction that means wasting some valuable information and poor classification results. Thus, we should pay more attention to multi-features. And on the other hand, due to extreme requirements for classification accuracy, we should hierarchically explore more deep features. The first thought is machine learning, but the traditional machine learning classifiers, like the support vector machine, are not friendly to larger inputs and features. This paper introduces a hybrid of principle component analysis (PCA), guided filtering, deep learning architecture into hyperspectral data classification. In detail, as a mature dimension reduction architecture, PCA is capable of reducing the redundancy of hyperspectral information. In addition, guided filtering provides a passage to spatial-dominated information concisely and effectively. According to the stacked autoencoders which is a efficient deep learning architecture, deep-level multi-features are not in mystery. Two public data set PaviaU and Salinas are used to test the proposed algorithm. Experimental results demonstrate that the proposed spectral–spatial hyperspectral image classification method can show competitive performance. Multi-feature learning based on deep learning exhibits a great potential on the classification of hyperspectral images. When the number of samples is 30 % and the iteration number is over 1000, the accuracy rates for both of the two data set are over 99 %.