Enhancing Hyperspectral Image Compression Through Stacked Autoencoder Approach

Hyperspectral Image (HSI) consisting of numerous high-resolution spectral bands creates challenges of the high dimensionality problem in HSI classification which hinder real-life applications despite their abundance of information. We present a lossy compression approach using stacked autoencoders to reduce high dimensionality problem in this paper. The pro-posed method utilizes stacked autoencoders to extract features from HSIs, allowing compression and subsequent reconstruction. The study demonstrates improved Peak Signal-to-Noise Ratio (PSNR) of 70.43, 60.87 and 61.36 on three distinct HSI datasets-Salinas, Botswana, and KSC respectively compared to previous autoencoder-based compression method. Additionally, the impact of compression on classification accuracy is assessed using a 3D-2D CNN model, achieving an average accuracy of 99 % across the datasets. Improved compression along with high classification accuracy shows our proposed method's usefulness.