Houénafa et al. (2026) Enhancing Conceptual Rainfall-Runoff Modeling in Data-Scarce Catchments using Machine Learning: Kolmogorov-Arnold Networks Compared to LSTMs

Identification

• Journal: Earth Systems and Environment
• Year: 2026
• Date: 2026-02-04
• Authors: Sianou Ezéckiel Houénafa, Mouhamadou Bamba Sylla
• DOI: 10.1007/s41748-025-01010-5

Research Groups

• Department of Mathematics, Pan African University Institute for Basic Sciences, Technology and Innovation, Nairobi, Kenya
• Research and Innovation Centre, African Institute for Mathematical Sciences, Kigali, Rwanda

Short Summary

This study evaluates the effectiveness of Kolmogorov-Arnold Networks (KANs) in enhancing conceptual rainfall-runoff modeling in data-scarce catchments using a two-stage error-correction approach. It finds that KAN-based hybrid models, particularly when combined with wavelet transform preprocessing, generally outperform LSTM-based and standalone models, especially for high-flow predictions.

Objective

• To investigate the effectiveness of Kolmogorov-Arnold Networks (KANs) and their wavelet-based variants (WKANs) in enhancing simulations within the GR6J conceptual rainfall-runoff model under data-scarce conditions.

Study Configuration

• Spatial Scale: Two data-scarce river basins in Sub-Saharan Africa: the Ouémé at Savè basin, Benin (approximately 23,600 km²), and the Yala basin, Kenya (3,351 km²).
• Temporal Scale: Daily time series data for precipitation, potential evapotranspiration, and streamflow. The Ouémé at Savè basin data spans mid-2002 to 2007, and the Yala basin data spans 2014-2019. Model performance was evaluated for daily and total monthly runoff.

Methodology and Data

• Models used: GR6J (Génie Rural à 6 paramètres Journalier) conceptual rainfall-runoff model; Kolmogorov-Arnold Networks (KANs); Long Short-Term Memory (LSTMs); Hybrid models (GR6J-KAN, GR6J-LSTM, GR6J-WKAN, GR6J-WLSTM). Discrete Wavelet Transform (DWT) was used for preprocessing, employing Daubechies (db1, db2) and Coiflet (coif1, coif2) mother wavelets.
• Data sources: Daily precipitation and potential evapotranspiration from Météo Benin (Ouémé at Savè) and Climate Engine (Yala). Daily streamflow measurements from the National Directorate of Water (DG-Eau), Benin (Ouémé at Savè) and Kenyan Water Resources Authority (WRA) (Yala).

Main Results

• Standalone KAN and LSTM models underperformed the GR6J conceptual model in both river basins.
• Wavelet transform preprocessing significantly improved the performance of standalone LSTM and KAN models, although their accuracy generally remained below GR6J.
• Hybrid models (GR6J-WKAN, GR6J-WLSTM, GR6J-KAN, GR6J-LSTM) consistently outperformed all standalone models in overall and total monthly runoff simulation accuracy.
• The GR6J-WKAN model achieved the highest Nash-Sutcliffe Efficiency (NSE) of 0.93 in the Ouémé at Savè basin (compared to 0.86 for GR6J alone) and 0.76 in the Yala basin (compared to 0.68 for GR6J alone).
• KAN-based hybrid models (GR6J-KAN, GR6J-WKAN) demonstrated stronger performance than their LSTM-based counterparts for high-flow simulation in both basins.
• Wavelet-based hybrid models consistently outperformed their non-wavelet counterparts.
• The benefit of lagged predictors was found to be catchment- and model-dependent, not universally improving performance in wavelet-based hybrid GR6J-ML models for the Savè catchment.

Contributions

• First application of Kolmogorov-Arnold Networks (KANs) and their wavelet-based variants (WKANs) within a hybrid framework to enhance conceptual rainfall-runoff modeling (GR6J) in data-scarce catchments.
• Demonstrates that KAN-based hybrid models, especially with wavelet preprocessing, significantly improve runoff prediction accuracy, particularly for high flows, compared to LSTM-based and standalone models.
• Proposes a novel two-stage error-correction approach where KANs model the residual errors of a conceptual hydrological model.
• Offers insights into the effectiveness of wavelet transform preprocessing in KAN-based hydrological models and the context-dependency of lagged input benefits.
• Releases open-source code to support reproducibility and further research.

Funding

• The study did not receive any external funding.

Citation

@article{Houénafa2026Enhancing,
  author = {Houénafa, Sianou Ezéckiel and Sylla, Mouhamadou Bamba},
  title = {Enhancing Conceptual Rainfall-Runoff Modeling in Data-Scarce Catchments using Machine Learning: Kolmogorov-Arnold Networks Compared to LSTMs},
  journal = {Earth Systems and Environment},
  year = {2026},
  doi = {10.1007/s41748-025-01010-5},
  url = {https://doi.org/10.1007/s41748-025-01010-5}
}