Adounkpe et al. (2025) Deep Learning at Two Timescales: Dual Neural Networks for Predicting Fast Urban and Slow Karst Floods

Identification

• Journal: Inżynieria Mineralna
• Year: 2025
• Date: 2025-11-05
• Authors: Julien Yise Peniel Adounkpe, Valentin Wendling, Alain Dezetter, Bruno Arfib, Kalil Traoré, Guillaume Artigue, Octavian Dobricean, Séverin Pistre, Anne Johannet
• DOI: 10.29227/im-2025-02-03-27

Research Groups

• HSM (Univ. Montpellier, IMT Mines Ales, IRD, CNRS)
• HydroSciences Montpellier, Univ. Montpellier, IRD, CNRS, UFR Pharmacie
• Aix Marseille Univ., CNRS, IRD, INRAE, CEREGE
• Synapse Informatique

Short Summary

This study develops and evaluates dual artificial neural networks (ANNs) for predicting fast urban and slow karst flash floods in the Las River, France, demonstrating that combining specialized ANNs offers the most robust and generalizable flood forecasting performance across both hydrological regimes.

Objective

• To develop and compare ensemble artificial neural network (ANN) strategies, specifically multilayer perceptrons (MLPs), for predicting flash floods in the Las River, Toulon (France), which exhibits a complex dual hydrological regime of rapid urban runoff and slower karst groundwater response.

Study Configuration

• Spatial Scale: Las River basin in Toulon, south-east of France, encompassing an urban area of approximately 12 square kilometers and a karst impluvium of 70 square kilometers.
• Temporal Scale: Hydrometeorological database spanning 6 years (October 2012 to September 2018) with a data timestep of 15 minutes. Forecasting horizon of 30 minutes (two time-steps).

Methodology and Data

• Models used:
  • Artificial Neural Networks (ANNs), specifically Multilayer Perceptrons (MLPs).
  • Specialized models: Urban Runoff Model (UM), Karst Model (KM).
  • Ensemble models: Output Combination Model (OM), Structure Combination Model (SM), Bulk Model (BM).
  • Levenberg Marquardt (LM) algorithm for model training.
• Data sources:
  • Water level measurements from Ragas and Saint-Antoine springs and the Las River at Lagoubran.
  • Rainfall data from seven stations, downloaded from the French meteorological service website (https://meteo.data.gouv.fr/).

Main Results

• Specialized models (UM for urban runoff, KM for karst events) achieved the highest performance on their respective event types (UM PERS PEAK: 0.596 for urban; KM PERS: 0.369 for karst) but demonstrated poor generalization to other event types.
• The Output Combination Model (OM) proved to be the most robust ensemble strategy, exhibiting consistent accuracy across all event types (overall PERS PEAK: 0.573) and performing best on extreme test events.
• The Bulk Model (BM) showed strong performance on urban events (PERS PEAK: 0.520) and better persistence on karst events compared to the SM, ranking second among ensemble models.
• The Structure Combination Model (SM) was the least accurate among the ensemble modeling approaches.
• Input variable selection was found to have a greater impact on model performance than the complexity of the model's structure or hyperparameter tuning.
• The 30-minute lead time for karst events resulted in lower persistence scores, suggesting that models over-relied on groundwater levels rather than capturing more complex hydrological patterns for longer-term karst forecasts.

Contributions

• Proposes and evaluates an ensemble modeling strategy using dual Artificial Neural Networks (MLPs) to effectively predict flash floods in complex hydrosystems characterized by distinct fast (urban runoff) and slow (karst groundwater) hydrological responses.
• Demonstrates the superior robustness and generalization capacity of combining specialized ANNs (Output Combination Model) for dual-regime flood forecasting compared to single "bulk" models or structural combinations.
• Highlights the critical importance of input variable selection over model structure complexity for improving prediction accuracy in complex hydrological systems.
• Addresses the challenge of forecasting with a short lead time (30 minutes) for both fast urban and slow karst events, identifying limitations for karst events and suggesting future improvements for operational conditions.

Funding

• French ANR (ANR-21-LCV1-0001) for the Laboratory of Hydrological forecasting by Artificial Intelligence (Hydr.IA) project.

Citation

@article{Adounkpe2025Deep,
  author = {Adounkpe, Julien Yise Peniel and Wendling, Valentin and Dezetter, Alain and Arfib, Bruno and Traoré, Kalil and Artigue, Guillaume and Dobricean, Octavian and Pistre, Séverin and Johannet, Anne},
  title = {Deep Learning at Two Timescales: Dual Neural Networks for Predicting Fast Urban and Slow Karst Floods},
  journal = {Inżynieria Mineralna},
  year = {2025},
  doi = {10.29227/im-2025-02-03-27},
  url = {https://doi.org/10.29227/im-2025-02-03-27}
}

Generated by BiblioAssistant using gemini-2.5-flash (Google API)