Predicting Wildfire Propagation in Europe, the Middle East and North Africa using Convolutional Neural Networks with an adjusted Dataset

  • Wildfires pose a significant threat to ecology, economy, and human lives alike. Droughts and heat waves fueled many fire occurrences in the last years, and the ongoing climate change increases the risk of larger, more devastating events. As part of the TREEADS project, funded by Horizon 2020 (EU), we are developing a wildfire propagation model, which is an important part of possible counter measures to support decision makers and firefighters in their actions against uncontrolled fire spread. With recent advances, machine learning became applicable for wildfire propagation modelling. Once the time-consuming training process is finished, predictions are fast, even on devices with low computational power. On the downside, large datasets are crucial to train robust models, but temporally accurate propagation data of real fire occurrences are sparse. To solve this problem, we reconstructed wildfire propagation in 12-hour intervals for over 5500 events with varying sizes and durations inWildfires pose a significant threat to ecology, economy, and human lives alike. Droughts and heat waves fueled many fire occurrences in the last years, and the ongoing climate change increases the risk of larger, more devastating events. As part of the TREEADS project, funded by Horizon 2020 (EU), we are developing a wildfire propagation model, which is an important part of possible counter measures to support decision makers and firefighters in their actions against uncontrolled fire spread. With recent advances, machine learning became applicable for wildfire propagation modelling. Once the time-consuming training process is finished, predictions are fast, even on devices with low computational power. On the downside, large datasets are crucial to train robust models, but temporally accurate propagation data of real fire occurrences are sparse. To solve this problem, we reconstructed wildfire propagation in 12-hour intervals for over 5500 events with varying sizes and durations in Europe, the Middle East, and North Africa. Thereto, burned area polygons from the European Forest Fire Information System (EFFIS) database were combined with active fire detection points from the Visible Infrared Imaging Radiometer Suite (VIIRS). The fire spread was reconstructed sequentially according to the revisiting times of VIIRS. This data was coupled with meteorological information from the ERA5 reanalysis product and surface information derived from Sentinel-2, as well as TanDEM-X remote sensing data. The aggregated dataset was then used to build a deep-learning convolutional neural network that captures meteorological effects, elevation, and vegetation on wildfire propagation. To account for the sequentially updated weather data from ERA5, long short-term memory (LSTM) with self-attention was included. In summary, we have constructed a novel wildfire propagation dataset suitable for machine learning purposes and developed a convolutional LSTM network for rapid prediction of fire spread.zeige mehrzeige weniger

Volltext Dateien herunterladen

  • AGU24_SimonMueller_fin.pdf
    eng

Metadaten exportieren

Weitere Dienste

Suche bei Google Scholar Anzahl der Zugriffe auf dieses Dokument
Metadaten
Autor*innen:Simon Müller
Koautor*innen:Anja Hofmann-BöllinghausORCiD, Kristin Vogel, Philipp BennerORCiD
Dokumenttyp:Vortrag
Veröffentlichungsform:Präsentation
Sprache:Englisch
Jahr der Erstveröffentlichung:2024
Organisationseinheit der BAM:7 Bauwerkssicherheit
7 Bauwerkssicherheit / 7.5 Technische Eigenschaften von Polymerwerkstoffen
VP Vizepräsident
VP Vizepräsident / VP.1 eScience
DDC-Klassifikation:Technik, Medizin, angewandte Wissenschaften / Ingenieurwissenschaften / Sanitär- und Kommunaltechnik; Umwelttechnik
Freie Schlagwörter:Convolutional Neural Network (CNN); Deep Learning; Remote Sensing; Wildfire
Themenfelder/Aktivitätsfelder der BAM:Infrastruktur
Infrastruktur / Fire Science
Veranstaltung:AGU24
Veranstaltungsort:Washington D.C., USA
Beginndatum der Veranstaltung:09.12.2024
Enddatum der Veranstaltung:13.12.2024
Verfügbarkeit des Dokuments:Datei im Netzwerk der BAM verfügbar ("Closed Access")
Datum der Freischaltung:19.11.2025
Referierte Publikation:Nein
Eingeladener Vortrag (wissenschaftliche Konferenzen):Nein
Einverstanden
Diese Webseite verwendet technisch erforderliche Session-Cookies. Durch die weitere Nutzung der Webseite stimmen Sie diesem zu. Unsere Datenschutzerklärung finden Sie hier.