Comparison of 1D and 3D Convolutional Neural Networks for Wildfire Detection Using PRISMA Hyperspectral Imagery and Domain Adaptation

This research explores the potential use of artificial intelligence techniques and edge computing approaches to detect wildfires directly from satellite platforms. The study is based on PRISMA (Hyperspectral Precursor of the Application Mission), an Italian hyperspectral satellite launched in 2019 that provides hyperspectral imagery in the spectral range of 0.4-2.S
$\mu \mathrm{m}$
with an average spectral resolution of less than 10 nm. The paper presents new results related to the Australian fires that occurred in December 2019 in New South Wales, acquired by PRISMA on December 27, 2019. The paper aims to investigate the practicality of deploying a one and three-dimensional convolutional neural network (CNN) models, as previously proposed by previous authors' works, with the assistance of an Nvidia Jetson TX2 as a testing hardware accelerator. This experiment explores the potential of utilizing on-the-edge deployment for this technology. This study aligns with efforts to improve the computational capabilities and autonomy of satellites, which could pave the way for future satellites or constellations with a specific focus on remote sensing and the provision of timely and reliable alerts.