Transferability of Convolutional Autoencoder Model For Lossy Compression to Unknown Hyperspectral Prisma Data

This work addresses the challenge of the portability of Autoencoder models for the lossy compression of different spatially independent and unknown hyperspectral satellite data. We propose an advanced 1D-Convolutional Autoencoder architecture for lossy hyperspectral data compression with high transferability to unknown spectral signatures. In the first experiment, the model is trained on a single PRISMA data set, and in the second experiment it is trained on five PRISMA data sets from all over the world. The abstraction ability of the two models is verified by processing six spatially independent hyperspectral PRISMA satellite data sets. The evaluation is based on the reconstruction accuracy using the SNR and SA metrics and compares it to other learning-based lossy compression techniques. We demonstrate the high transferability and generalization of our 1D-Convolutional Autoencoder for a fixed compression ratio on each PRISMA satellite data set, which results in superior reconstruction accuracy compared to state-of-the-art methods.