Deep learning for downscaling GOES-18 measurements for wildfire detection

Abstract

This thesis aims to address the challenge of accurate wildfire detection using satellite imagery. Despite the availability of various satellite-based fire products, real-time detection of fire perimeters remain difficult due to limitations in the spatio-temporal resolution of current satellite imagery. For example, the Geostationary Operational Environmental Satellites (GOES-R) series containing the Advanced Baseline Imager (ABI) offers high temporal resolution for frequent observations but suffers from low spatial resolution. In contrast, low Earth orbit (LEO) satellites like Suomi-NPP, NOAA-20, and NOAA-21 with the Visible Infrared Imaging Radiometer Suite (VIIRS) imager provide high spatial resolution but with limited temporal coverage. To overcome these limitations, this research proposes a deep learning framework for wildfire detection that leverages GOES ABI observations, which are downscaled to a spatial resolution of 375 meters using a Generative Adversarial Network (GAN). High-resolution VIIRS images are used as ground truth labels during the training phase. Experimental results demonstrate that the proposed framework successfully enhances the spatial resolution of GOES ABI data while preserving its high temporal frequency, allowing more precise and timely wildfire detection.