Casas et al. (2026) Using machine learning for the prediction of flood-related 112 calls

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2026
• Date: 2026-08-20
• Authors: Jordi Morales Casas, Andreas Kaltenbrunner, Àgata Lapedriza, Xavier Llort
• DOI: 10.5194/nhess-26-3943-2026

Research Groups

• AIWELL Lab, Universitat Oberta de Catalunya, Rambla del Poblenou, 154, 08018 Barcelona, Spain
• Hydrometeorological Innovative Solutions (HYDS), Jordi Girona 1–3, ParcUPC K2M, 08034 Barcelona, Spain
• Department of Engineering, Universitat Pompeu Fabra, Carrer de Roc Boronat 138, 08018 Barcelona, Spain
• Institute for Experiential AI, Northeastern University, 360 Huntington Ave, Boston, MA 02115, USA

Short Summary

This study proposes an ML methodology for predicting flood-related impacts using emergency calls as the target variable. The model combines rainfall-related weather data, information on recent emergencies, and static vulnerability-exposure layers to predict flood-related impacts at a municipal and hourly resolution.

Objective

• Develop an ML-based prediction problem that integrates meteorological hazard data with information on vulnerability and exposure into data-driven forecasting systems.
• Evaluate the capacity of the methodology to increase detection power and reduce false alarms when compared to hazard-based baselines across different population densities.

Study Configuration

• Spatial Scale: Municipality-level resolution in Catalonia, northeastern Spain.
• Temporal Scale: Hourly temporal resolution between October 2018 and February 2025.

Methodology and Data

• Models used: Machine learning (ML) models, specifically random forest classifiers.
• Data sources:
  • Emergency call data from Catalan Civil Protection authorities.
  • Automatic weather station data from SMC's network of AWSs.
  • Official weather warnings from AEMET and SMC.
  • Radar data from SMC's network of C-band Doppler radars.

Main Results

• The ML approach represents a substantial improvement in two out of the three population density groups.
• The model for the lowest-density group struggles due to a lack of impact data, highlighting a key roadblock for data-driven algorithm development in sparsely populated regions.
• Feature importance analyses and ablation experiments show that the models can effectively combine various data sources to provide more effective predictions than using these sources in isolation.

Contributions

• The study contributes to the development of impact-based forecasting systems by integrating meteorological hazard data with information on vulnerability and exposure.
• The methodology addresses the severe temporal class imbalance and uncertainty characteristics of emergency call data, enabling the evaluation of performance across diverse population dynamics and varying data availability.

Funding

• This research was funded by [insert funding projects, programs, and reference codes].

Citation

@article{Casas2026Using,
  author = {Casas, Jordi Morales and Kaltenbrunner, Andreas and Lapedriza, Àgata and Llort, Xavier},
  title = {Using machine learning for the prediction of flood-related 112 calls},
  journal = {Natural hazards and earth system sciences},
  year = {2026},
  doi = {10.5194/nhess-26-3943-2026},
  url = {https://doi.org/10.5194/nhess-26-3943-2026}
}