Robust compression and classification of hyperspectral images

Abstract

Hyperspectral images are a powerful source of spectral data that have been utilized in a wide array of applications. The large size of hyperspectral images limits the applicable uses and necessitates effective compression methods. While many spectral-spatial compressors have been proposed in the past, there has been little work on the benefits of a spectral-only strategy. A spectral-only strategy not only has compressive capabilities but would also allow the classification of the compressed images, making the contributions of this thesis multi-fold. We present a Long Short Term Memory Autoencoder designed for the spectral compression of hyperspectral images. We show that this network can compress the images effectively with low reconstruction error, as well as require fewer training parameters to compress when compared to existing spectral-spatial compression methods. Existing learned compression models often require many of the pixels to be used to train an image. We demonstrate that our proposed network does not suffer a reduction in compression performance by reducing the number of training examples. Existing compression techniques are limited in capability by their inclusion of spatial information, requiring reprocessing for all images that have sufficiently different scenes. We demonstrate the proposed network's robustness by training a single model for use in multiple scenes without the requirement of retraining the model from scene to scene. Furthermore, using the feature extracting capabilities of an autoencoder, we analyzed the capabilities of the compressed image as a feature set for classification. Experimental results demonstrate that the unsupervised compressed features generated can be utilized for supervised machine learning classification tasks. We also demonstrate that the robustness of the compressor allowed for a single network to not require being retrained for compressing and then classifying new images without significant loss.