Synthetic generation of hyperspectral data from multispectral data through generative adversarial network model

Remote sensing applications utilizing hyperspectral imaging technology have demonstrated their capability to deliver detailed and comprehensive analytical results in earth observation studies, significantly outperforming traditional multispectral data approaches. One constraint of these methods, however, is the limited datasets. The PRISMA (PRecursore IperSpettrale della Missione Applicativa) mission, exemplifies the challenge of obtaining timely hyperspectral data with its low revisit time. Given these constraints, development and production of synthetic hyperspectral PRISMA imagery derived from more readily available multispectral data sources, such as Landsat 8/9, has emerged as a critical objective. The dimensionality of PRISMA data is too high for traditional extrapolation, therefore this work explores variations of feed forward networks (FFNs) to generate synthetic PRISMA images from an existing multispectral signal. Experiments are run with both a MLP (multi layer perceptron) approach and a specialized GAN (generative adversarial network) for image synthesis. The Optuna library and grid search algorithms are used for optimizing network hyperparameters. NDVI (normalized difference vegetation index) is used to validate the accuracy of the synthesized image against real data. The final training performance displays excellent generalization (0.84 coefficient of correlation across train, test and validation sets and a consistently > 0.7 coefficient of correlation across all external tests).