Deep Learning Approach to Improve Spatial Resolution of GOES-R Satellite Imagery for Active Fire Detection Using VIIRS Satellite Data

Deep Learning Approach to Improve Spatial Resolution of GOES-R Satellite Imagery for Active Fire Detection Using VIIRS Satellite Data Qunying Huang1, Haiyue Liu2, Tang Sui1, Songxi Yang1, Song Gao1 1Department of Geography, University of Wisconsin – Madison (UW–Madison) 2Department of Computer Science, UW–Madison Wildfires are increasingly frequent and destructive worldwide [1], posing severe risks to ecosystems [2-5], infrastructure [6-10], and public safety [11-13]. As global warming exacerbates these risks, highlighting the urgent need for advanced active wildfire detection models to support timely response and damage mitigation. Current active wildfire (AF) detection products, such as NOAA’s Geostationary Operational Environmental Satellite-R series (GOES-R) AF product [14], and NASA’s MCD14DL [15] and VNP14IMGTDL_NRT [16, 17] based on Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) imagery, respectively, enable large-scale monitoring of wildfire occurrences. However, these products are constrained by threshold-based methods, and outdated statistical or physical models, limiting their accuracy and generalizability across diverse geographical regions and fire conditions [18, 19]. In addition, existing active fire products face spatial-temporal trade-offs: GOES-R AF product provides high temporal (5 min for the Continental United States (CONUS)–15 min for the Western Hemisphere/full disk) but low spatial spatial resolution (2 km), resulting in missed detections of small or low-intensity fires, overestimation of fire size, and a high false alarm rate. In contrast, MODIS and VIIRS AF products offer finer spatial resolution (500 m, 375 m) but a 12hour revisit time, resulting in significant temporal gaps that restrict real-time applications. In response, this study proposes a two-step framework: (1) super-resolution (SR) of GOES-R data to approximate VIIRS-level detail, and (2) application of a transformer based deep learning (DL) model to detect active fires in super-resolved images. The approach exploits the complementary spatial and temporal strengths of both sensors, enabling detection of small, lowintensity, and fast-spreading fires often missed by threshold-based methods. While DL methods have been used separately for SR enhancement and active fire detection with MODIS [20] and VIIRS [21] imagery, limited studies have integrated both for near-real-time detection with GOESR data. Our super-resolution model follows a modern encoder–decoder autoencoder (AE; Figure 1d) with skip connections and transformer-style attention blocks that performs multi-scale upsampling on a co-registered GOES-R crop over the VIIRS footprint. The AE maps low-resolution GOES-R ABI image to a VIIRS-like, high-resolution data for downstream active-fire detection. Training uses near-coincident GOES-R and VIIRS event pairs for pixel-wise correspondence, with a physicsaware loss that enforces radiometric fidelity, structural preservation, downsample-consistency back to GOES-R scale, and hotspot-aware weighting for fire lines and active cores. The superresolved output, together with an uncertainty channel and the original GOES-R context, is then fed to a lightweight transformer-based segmentation model to produce per-pixel fire probabilities and crisp boundaries. Preliminary results indicate that the proposed AE-based SR model effectively sharpens and localizes the hotspot toward the VIIRS peak while suppressing background fluctuations (Figure 1c). In addition, the predicted fire mask (Figure 1e) exhibits cleaner and more compact detections from the super-resolved GOES-R imagery compared with the GOES-R derived baseline, supporting GIS-ready polygon extraction. This study demonstrates the feasibility of superresolved GOES-R imagery for real-world active fire detection, offering practical contributions to early warning systems that support emergency response and resource allocation. Figure 1. Downscale GOES-R thermal band 7 (a) to approximate VIIRS thermal band I4 (b) for active fire detection with an autoencoder architecture (d); The model sharpens and localizes the hotspot toward the VIIRS peak (c) while suppressing background fluctuations; The predicted fire mask (e right) row demonstrates cleaner, more compact detections from the model output compared with a GOES-derived baseline (e left). Reference 1. Jones, M.W., et al., State of wildfires 2023–2024. Earth System Science Data, 2024. 16(8): p. 36013685. 2. Gajendiran, K., S. Kandasamy, and M. Narayanan, Influences of wildfire on the forest ecosystem and climate change: A comprehensive study. Environmental Research, 2024. 240: p. 117537. 3. Wardle, D.A., et al., Long-term effects of wildfire on ecosystem properties across an island area gradient. Science, 2003. 300(5621): p. 972-975. 4. Lucas‐Borja, M.E., et al., Changes in ecosystem properties after post ‐ fire management strategies in wildfire ‐ affected Mediterranean forests. Journal of Applied Ecology, 2021. 58(4): p. 836-846. 5. Madani, N., et al., The impacts of climate and wildfire on ecosystem gross primary productivity in Alaska. Journal of Geophysical Research: Biogeosciences, 2021. 126(6): p. e2020JG006078. 6. Schulze, S.S., et al., Wildfire impacts on schools and hospitals following the 2018 California Camp Fire. Natural Hazards, 2020. 104: p. 901-925. 7. Fraser, A.M., M.V. Chester, and B.S. Underwood, Wildfire risk, post-fire debris flows, and transportation infrastructure vulnerability. Sustainable and Resilient Infrastructure, 2022. 7(3): p. 188-200. 8. Naser, M. and V. Kodur, Vulnerability of structures and infrastructure to wildfires: a perspective into assessment and mitigation strategies. Natural Hazards, 2025: p. 1-21. 9. Papalou, A. and D.K. Baros, Assessing Structural Damage after a Severe Wildfire: A Case Study. Buildings, 2019. 9(7): p. 171. 10. Habermann, N. and R. Hedel, Damage functions for transport infrastructure. International journal of disaster resilience in the built environment, 2018. 9(4/5): p. 420-434. 11. Paveglio, T., M.S. Carroll, and P.J. Jakes, Alternatives to evacuation—Protecting public safety during wildland fire. Journal of Forestry, 2008. 106(2): p. 65-70. 12. Udren, E.A., et al., Managing wildfire risks: Protection system technical developments combined with operational advances to improve public safety. IEEE Power and Energy Magazine, 2022. 20(1): p. 64-77. 13. Penney, G., et al., A fire safety engineering approach to improving community resilience to the impacts of wildfire. Fire and Materials, 2024. 14. NGFS. Next Generation Fire System 2025; Available from: https://cimss.ssec.wisc.edu/ngfs/. 15. Team, N.V.L.S., VIIRS (NOAA-21/JPSS-2) I Band 375 m Active Fire Product NRT (Vector data), N.L.M.a.t. MODAPS, Editor. 2021. 16. Jiao, M., Z. Kang, and X. Quan. Evaluation of Fire Products Using Spatio-Temporal Clustering Method. in IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium. 2022. IEEE. 17. Csiszar, I., et al., Active fires from the Suomi NPP Visible Infrared Imaging Radiometer Suite: Product status and first evaluation results. Journal of Geophysical Research: Atmospheres, 2014. 119(2): p. 803-816. 18. Giglio, L., et al., The Collection 6 MODIS burned area mapping algorithm and product. Remote sensing of environment, 2018. 217: p. 72-85. 19. Boschetti, L., et al., Global validation of the collection 6 MODIS burned area product. Remote sensing of environment, 2019. 235: p. 111490. 20. Giglio, L., W. Schroeder, and C.O. Justice, The collection 6 MODIS active fire detection algorithm and fire products. Remote sensing of environment, 2016. 178: p. 31-41. 21. Schroeder, W., et al., The New VIIRS 375 m active fire detection data product: Algorithm description and initial assessment. Remote Sensing of Environment, 2014. 143: p. 85-96.