Transfer Learning Analysis For Wildfire Segmentation Using Prisma Hyperspectral Imagery And Convolutional Neural Networks

• Published in: 2022 12th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS) pp. 1 - 5
• Main Authors: Spiller, Dario, Amici, Stefania, Ansalone, Luigi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 13.09.2022
• Subjects: Convolutional Neural Networks; Fires; Hyperspectral Imagery; Image resolution; Image segmentation; PRISMA; Signal processing; Space missions; Training; Transfer learning; Wildfires
• ISSN: 2158-6276
• DOI: 10.1109/WHISPERS56178.2022.9955054

In this work we present a segmentation study of wildfire scenarios using PRISMA hyperspectral data and a methodology based on convolutional neural networks and transfer learning. PRISMA (Precursore IperSpettrale della Missione Applicativa, Hyperspectral Precursor of the Application Mission) is the hyperspectral mission by ASI (Agenzia Spaziale Italiana, Italian Space Agency) launched in 2019 providing images with a spectral range of 0.4-2.5 \mu \text{m} and an average spectral resolution less than 10 nm. We used the PRISMA hypercube acquired during the Australian bushfires of December 2019 in New South Wales to train a one-dimensional convolutional neural network and perform a transfer learning in the Bootleg Fire of July 2021 in the Fremont-Winema National Forest in Oregon. The generalization ability of the network is discussed and potential future applications are presented.