Transfer Learning for Hyperspectral Images Utilizing Channel Selection Techniques and Ensemble Methods

Abstract

Hyperspectral images contain information from a wider range of the electromagnetic spectrum than natural images which gives them potential for better classification ability. However, hyperspectral datasets are typically small due to the expensive equipment needed to obtain the images, which can limit classification performance. One solution to this problem is transfer learning, in which a model trained on one dataset is reused for a separate dataset. Research has shown that transfer learning between hyperspectral datasets can give improved performance over models without transfer learning when training data are limited. Since extra hyperspectral data are not always available, the solution proposed here is to instead use networks pretrained on natural image (i.e., red, blue, green, or RGB) datasets for transfer learning. By using various feature selection and feature extraction methods, extracted hyperspectral samples are transformed into a three-channel format to imitate an RGB image and are used for fine tuning the well-known ResNet, DenseNet, and VGG networks. Feature extraction methods include techniques like principal component analysis, which create lower dimensional features from high dimensional spectral data. Alternatively, feature selection methods aim to find the best set of existing channels to use for classification. Experimental results are obtained using two well-known hyperspectral datasets, showing 73.6% accuracy on Pavia University and 82.8% accuracy on Salinas with 25 training samples per class. Additional ensemble methods are implemented that utilize multiple networks and show an increase in accuracy of 4.4% and 3% for Pavia University and Salinas, respectively. These results demonstrate that networks pretrained on RGB datasets are suitable for transfer learning with hyperspectral image datasets and can achieve desirable performance given the proper preprocessing technique.