Hyperspectral Image Compression using Modified Convolutional Autoencoder

Abstract

Spectral imaging is a type of multi band imaging technique of the electromagnetic spectrum, used for gathering and analysis of information. Hyperspectral imaging is a technique that collects spectral information from a broad spectrum of wavelengths for the same spatial area of each pixel. Due to its multiple bands, and spectral and spatial redundancy, the image size is immense. Processing these images requires an enormous amount of memory. Our paper proposes a lossy technique to encode and compress the hyperspectral images by the help of deep convolutional networks, autoencoders and attention layers. The encoder uses convolutional and max pooling layers, connected to a singular attention layer to encode the data, and the decoder has a single dense layer preceding transpose convolutional and upsampling layers to decode the coded data back into the hyperspectral images. The method is tested on different images taken by the Airborne Visible/ Infrared Imaging Spectrometer (AVIRIS), Reflective Optics System Imaging Spectrometer (ROSIS), and NASA EO1 satellite. The method achieves superior results than existing work by up to a 5% increase in the PSNR and up to 200 times increase in the compression ratio.