GIDPC.Green Infrastructure · Disaster Prevention
Research & Technology
Advances in Water ResourcesUnder Review

Continuous Soil Moisture Dynamics for Urban Stormwater Modeling Using Neural ODE: A Physics-Informed Approach for Seoul

Shows that the static AMC framework of EPA-SWMM cannot represent soil-moisture variability in East-Asian monsoon climates, and replaces it with a physics-informed Neural ODE that learns continuous dθ/dt with R² = 0.983.

Authors
Junsuk Kang
Year
2026

Keywords

soil moisture dynamicsNeural ODEurban hydrologyGreen-Ampt infiltrationantecedent moisture conditionphysics-informed machine learning

Background

Urban hydrology models such as EPA SWMM still represent antecedent moisture condition (AMC) as a discrete three-class proxy. Soil moisture, however, is inherently continuous, and forcing it into three bins severely degrades LID-evaluation credibility in monsoon climates.

For Seocho-gu (2018–2023), ERA5-Land data for 21 observed storms show that the canonical AMC II value (θ₀ = 0.170 m³/m³) underestimates observed moisture in every single event, with mean bias 0.224 m³/m³ and RMSE 0.238 m³/m³.

Method

Two coupled modules:

  1. GE-Water — a new distributed urban hydrology engine combining Green-Ampt infiltration with 2D diffusion-wave surface routing on a 50 m grid; continuous θ(t, y, x) tracking; generates training data without proprietary models.
  2. Physics-informed Neural ODE — a neural network learns dθ/dt as a function of antecedent state and real-time rainfall.

Trained on 84 synthetic simulations (21 events × 4 initial moisture conditions) in 95 seconds.

Key Results

  • Validation R² = 0.983, MAE = 0.56% (of mean θ)
  • Per-event-size MAE 0.0003–0.0060 m³/m³, uniform performance across the full antecedent spectrum
  • All 21 events exceed the AMC III boundary — calling for structural redesign, not simple recalibration

Significance

A computationally efficient (95-second training), physically interpretable pathway to continuous soil-moisture initialization for urban stormwater and LID performance modeling.