DEWMOST (Deep-learning and Ensemble Watershed Modeling Of Stream Turbidity) is a hybrid modeling framework developed to address the increasing variability of sediment runoff caused by extreme rainfall events and climate change. By integrating traditional deterministic watershed models with advanced deep learning architectures, DEWMOST overcomes the limitations of single-model approaches, providing highly accurate and stable predictions of stream turbidity and sediment transport even under extreme hydrological conditions.

• Soyang Lake Watershed / Korea Environmental Industry & Technology Institute (2022 – 2023)

• Paldang Lake Watershed / Korea Environmental Industry & Technology Institute (2024 – 2025)

DEWMOST operates through a sophisticated integration of optimized ensemble processes and AI-driven post-processing:

• Statistically Optimized Ensemble Operation:

  • To ensure representativeness while excluding redundant information, the system employs Statistical Clustering to select a concise set of representative models.

  • These models are chosen from a pool of 18 rainfall-runoff models and 72 sediment transport models, customized to reflect specific regional characteristics.

• Transformer-based Time-series Prediction (Meta-Learner):

  • Contextual Sequence Learning: Directly learns the temporal relationships and time-lag effects from historical data (120 hours of runoff and 5 days of sediment concentration).

  • Self-Attention Mechanism: Dynamically weighs the correlations between input precipitation and model-predicted values, focusing on the most critical information to derive the final output.

• Flow-Conditioned Hybrid Prediction (MoE Architecture):

  • Recognizes that sediment behavior differs fundamentally between low-flow (baseflow) and high-flow (flood) conditions.

  • Implements a Mixture of Experts (MoE) system, consisting of dual specialized models for low and high-concentration scenarios.

  • Gating Network Integration: A dedicated Gate Network evaluates the current hydrological status and adaptively fuses the outputs of the expert models to calculate the optimal final sediment concentration.