Flood Risk Mapping Through Advanced Machine Learning Techniques and Geomorphic Data Integration
- 1Dipartimento di Ingegneria Civile, Edile e Ambientale, Università degli Studi di Napoli Federico II - Via Claudio, 21, 80125 Napoli, Italia. (jorgeandres.saavedranavarro@unina.it)
- 2Dipartimento di Scienze del Suolo, della Pianta e degli Alimenti, Università degli Studi di Bari Aldo Moro - Via Giovanni Amendola, 165/a, 70126 Bari, Italia. (cinzia.albertini@uniba.it)
Flood events rank among the most destructive natural hazards, necessitating comprehensive risk management strategies to mitigate their impact on human health, the environment, cultural heritage, and economic activities. In this context, various approaches have been developed for identifying flood-prone areas, but there is still a need to enhance their capabilities due to dynamic changes in landscape and infrastructure.
In recent years, there has been a proliferation of remote sensing observations that can support dynamic and continuous mapping of flood-prone areas by integrating the most updated information. This study explores the potential of machine learning (ML) techniques, including Random Forest, Support Vector Machine, and Navies Bayer model, utilizing geomorphic information such as slope, elevation, precipitation, land use/land cover, elevation difference to the nearest river, and others as predictor variables. The best model and set of variables were explored by adopting approximately 30 variables spanning types, hydrologic, topographic, and categorical categories. Careful consideration was given to avoiding high correlations between variables in test subsets, ensuring relevance, and avoiding redundancy. Calibration and validation of the model employ Copernicus Emergency Management Service maps from Sentinel-2 satellite coupled with regional maps of past flood events.
Results highlight that the best ML technique is represented by the Random Forest, adopting a range of 5 to 8 variables for effective delineation of flood-prone areas. Among the selected variables, the most relevant ones include Rainfall, Geomorphic Flood index - GFI, Lithology, and others. The study demonstrates that a minimal amount of information (between 0.1% and 10%) suffices for optimal model performance (AUC greater than 0.8).
The study covered the entire territory of Italy, resulting in a flood-prone map at a 90m resolution, validated with flood maps provided by national agencies and obtained through traditional hydraulic models.
Keywords: satellite images, flood-prone areas, Machine Learning, GFI, flood risk.
How to cite: Saavedra Navarro, J., Zhuang, R., Albertini, C., and Manfreda, S.: Flood Risk Mapping Through Advanced Machine Learning Techniques and Geomorphic Data Integration, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17663, https://doi.org/10.5194/egusphere-egu24-17663, 2024.
