Top

WaveletKAN-GRU Model for Mountain Flood Sediment Yield Estimation Based on Rainfall-runoff and Terrain Data

Authors

Thanos Hadjis
Department of Data Science and Machine Learning, School of Electrical and Computer Engineering, National Technical University of Athens, Athens, 15780, Greece
Ilias Matsas
Department of Data Science and Machine Learning, School of Electrical and Computer Engineering, National Technical University of Athens, Athens, 15780, Greece

Files

Abstract

Short-duration Mountain floods can mobilize a large amount of sediment in a short time, but estimating the released sediment is still difficult because rainfall intensity, runoff concentration, slope connectivity, antecedent wetness and channel storage all interact nonlinearly. Propose a WaveletKAN-GRU model for mountain-flood sediment yield prediction based on rainfall-runoff and terrain data in this paper. Rainfall intensity, cumulative rainfall, antecedent moisture, runoff depth, discharge change rate, slope, relief, drainage density, topographic wetness index, land-cover stability and channel-connectivity indicators are all aligned into event windows. Wavelet decomposition is used to separate short storm pulses, recession components and low-frequency catchment memory, and thus abrupt rainfall-runoff changes are not smoothed out in the hydrograph. The KAN layer learns non-linear terrain-conditioned hydrological effects and shows that the same runoff signal can result in different sediment responses under different slope, wetness and channel-connectivity conditions. The GRU module then learns the delayed response of sediment in the event window, such as sediment release in rising limbs and continuous transport during recession. A simulated multi-watershed dataset with event-separated training, validation and test sets is used for the experiment. The model is compared to tree- based, convolutional, recurrent and wavelet-recurrent baselines. The results are shown as sediment-yield error, high-risk warning recall, confidence calibration and cross-event robustness. The output is the estimated sediment yield, confidence, warning level, dominant evidence and suggested watershed review action. The proposed framework is an explainable and mechanism-aware tool for the short-term sediment risk assessment of mountain watersheds.

Details