Baste et al. (2026) The need for uncertainty: why probabilistic LSTMs are key to improving flood predictions and enabling learned warning rules

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-10-16
• Authors: Sanika Baste, Sebastian Lerch, Daniel Klotz, Ralf Loritz
• DOI: 10.5194/hess-30-6339-2026

Research Groups

• Institute of Water and Environment, Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany
• Department of Mathematics and Computer Science, Marburg University, Marburg, Germany
• Machine Learning in Earth Science, Interdisciplinary Transformation University Austria, Linz, Austria

Short Summary

This study investigates whether probabilistic Long Short-Term Memory (LSTM) models improve the representation of extreme events in rainfall–runoff simulations and if reinforcement learning can translate these into actionable flood warnings. It finds that probabilistic LSTMs significantly enhance the capture of extreme flood events compared to deterministic LSTMs, and a reinforcement learning agent (FRiCA) effectively learns risk-aware flood warning rules, outperforming static heuristics.

Objective

• Which of the investigated probabilistic LSTMs provides the most reliable rainfall–runoff simulation?
• Why are probabilistic LSTMs more suitable to represent hydrological extremes and the tail regions of the marginal discharge distribution?
• Can reinforcement learning aid operational flood-warning communication by learning a risk-aware translation of predictive distributions into actionable warnings based on flood-intensity categories?

Study Configuration

• Spatial Scale: 196 catchments in Switzerland, derived from the CAMELS-CH dataset.
• Temporal Scale: Daily hydro-meteorological time series from 1 January 1981 to 31 December 2020. Models were trained from 1 October 1980 to 30 September 1995, validated from 1 October 1995 to 30 September 2000, and tested from 1 October 2000 to 30 September 2020.

Methodology and Data

• Models used:
  • LSTMdet (Deterministic Long Short-Term Memory network)
  • CMAL (Countable Mixture of Asymmetric Laplace Distributions) - probabilistic LSTM
  • DRN (Distributional Regression Network) - probabilistic LSTM using a logistic distribution
  • BQN (Bernstein Quantile Network) - nonparametric probabilistic LSTM
  • FRiCA (Flood Risk Communication Agent) - reinforcement learning agent based on Proximal Policy Optimization (PPO)
• Data sources:
  • CAMELS-CH dataset (Höge et al., 2023) providing daily hydro-meteorological time series and 22 static catchment attributes.
  • Meteorological forcings: precipitation (mm d⁻¹), minimum temperature (°C), maximum temperature (°C), and relative sunshine duration (%).
  • Target variable: specific discharge (mm d⁻¹).
  • Design floods with annual return intervals (2, 10, 30, 100, and 300 years) from BAFU (2024) for flood severity classification.

Main Results

• Deterministic LSTMs systematically underestimate 97% of extreme discharge events (top 0.05% of the marginal discharge distribution), with a mean underestimation of approximately 28 mm d⁻¹.
• The probabilistic CMAL model captures 83% of these extreme events within its 99th percentile prediction interval, demonstrating a significant improvement in representing flood extremes. DRN and BQN capture 60% and 57% respectively within their 99th percentile.
• The CMAL exhibits the lowest mean Continuous Ranked Probability Score (CRPS) of 0.50 and a median CRPS of 0.39, indicating superior overall calibration and sharpness among the probabilistic models, though its distributions are generally over-dispersed.
• The Flood Risk Communication Agent (FRiCA) achieves a higher hit count for floods (414) compared to the deterministic LSTM (301) and a static 99th percentile rule (124), while maintaining a more favorable false alarm rate than the static 99th percentile rule. For extreme floods, FRiCA also shows a higher hit count (27) than the deterministic LSTM (18).
• Probabilistic LSTMs, by estimating the full conditional discharge distribution, provide a more principled and interpretable representation of uncertainty under extreme forcing conditions, which is crucial for flood risk assessment.

Contributions

• Provides a systematic comparison of probabilistic LSTM architectures for rainfall–runoff simulation in Switzerland, specifically evaluating their ability to represent hydrological extremes.
• Demonstrates that the primary limitation in extreme flood prediction for deep learning models lies in the inherent uncertainty of input-target dynamics in regional hydrological modeling, which probabilistic models address more effectively than deterministic ones.
• Introduces and validates a novel reinforcement learning-based agent (FRiCA) for translating probabilistic flood forecasts into actionable, risk-aware categorical flood warnings, offering a dynamic alternative to static decision rules in early warning systems.
• Highlights the potential of reinforcement learning as an adaptable decision layer for flood risk communication, capable of learning user-defined risk preferences.

Funding

• Volkswagen Foundation (project "Invigorating Hydrological Science and Teaching: Merging Key Legacies with New Concepts and Paradigms" (ViTamins), reference code: 0069685)
• Deutsche Forschungsgemeinschaft (grant no. 521210228)
• Karlsruhe Institute of Technology (KIT) (for article processing charges)

Citation

@article{Baste2026need,
  author = {Baste, Sanika and Lerch, Sebastian and Klotz, Daniel and Loritz, Ralf},
  title = {The need for uncertainty: why probabilistic LSTMs are key to improving flood predictions and enabling learned warning rules},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-6339-2026},
  url = {https://doi.org/10.5194/hess-30-6339-2026}
}