Synthesis of PRISMA data from Landsat 8 or 9 data using machine learning tools

Abstract

The PRISMA mission, operated by Italian Space Agency (ASI), provides a relatively new hyperspectral imaging sensor that captures a wide range of narrow spectral bands. With the potential to acquire detailed hyperspectral images, increasing the availability of such data could significantly benefit various environmental monitoring fields, including forestry, agriculture, and climatology. This study aims to exploit multispectral Landsat 8/9 data, which are temporally and spatially aligned with PRISMA data, to train a Neural Network (NN) model that will then be used to generate synthetic hyperspectral PRISMA data from multispectral Landsat 8/9 images. The focus is on reflectance in the VNIR range of the spectrum, with Landsat 8/9 providing 5 bands and PRISMA offering more than 50 bands. The approach explores a traditional Deep Learning technique, where PRISMA synthetic data is generated after training a NN, specifically a Multi-Layer Sequential Model with optimized parameters to directly predict PRISMA band values from the corresponding Landsat 8/9 input. Additionally, the study explores the use of advanced technology as Generative Adversarial Networks (GANs) to simulate spectral band values. Best results are found with GAN models with training R2 score around 0.8, while test scores fluctuate between 0.68-0.81. The methodology outlined in this work can serve as a benchmark for future research exploring also alternative techniques for generating synthetic hyperspectral data.