Classification of Multi-Resolution Hyperspectral Data by Convolutional Neural Networks

To realize the pixel-based classification of hyperspectral data, wavelet transform is applied to produce two-dimensional data, which makes it possible to combine convolutional neural networks (CNNs). Although wavelet transform is usually used for time series data, it can also be applied to hyperspectral data which consist of many bands. By analyzing hyperspectral data with multiple resolutions, the characteristics of the spectrum can be captured on various scales. The features generated by the wavelet transform are used as input data to the CNN classifier to identify class-specific patterns. We verified this approach using two data sets and confirmed its effectiveness. Futhermore, the proposed method performed well even with a small number of training samples.