LITF-PA-2026-014 · UrbanTech / Climate Resilience

System and Method for Real-Time Urban Flash Flood Depth Prediction and Dynamic Pedestrian-Vehicle Routing Using Distributed Stormwater Level Sensor Networks, LiDAR-Derived Digital Elevation Models, and Physics-Constrained Graph Neural Networks

Defensive Prior Art Disclosure
Urban street during heavy rainstorm with IoT water level sensors in storm drains and digital flood depth prediction overlay
⚖️ Prior Art Notice: This document is published as defensive prior art under 35 U.S.C. § 102(a)(1). The inventions described herein are dedicated to the public domain as of the publication date above. This disclosure is intended to prevent the patenting of these concepts by any party.

Abstract

Disclosed is a system and method for predicting urban flash flood inundation depth at street-segment resolution and dynamically rerouting pedestrians and vehicles around dangerous flood zones before surface water reaches hazardous levels. The system deploys low-cost ultrasonic water level sensors (unit cost $18–35) inside existing storm drain inlets, catch basins, and manholes across a municipal drainage network. Each sensor node measures water level at 10-second intervals and transmits readings via LoRa mesh to a central inference server. The server ingests these real-time hydraulic observations alongside rainfall intensity data from weather radar (NEXRAD Level III, 250 m resolution, 5-minute updates) and pre-computed drainage network topology extracted from municipal GIS databases. A physics-constrained graph neural network (PC-GNN) models the stormwater drainage network as a directed graph where nodes represent junctions (manholes, catch basins, outfalls) and edges represent pipe segments with known diameter, slope, and Manning's roughness coefficient. The GNN's message-passing layers are constrained by the Saint-Venant shallow water equations, ensuring predictions respect conservation of mass and momentum. When the model predicts that pipe capacity will be exceeded at any junction within the next 15–60 minutes, it computes surface overflow volumes and routes them across a LiDAR-derived digital elevation model (DEM, 1 m horizontal resolution) using a fast D8 flow accumulation algorithm to generate a street-level flood depth forecast map. This map feeds a routing API that issues real-time detour advisories to navigation applications, municipal alert systems, and connected vehicle platforms, steering traffic away from predicted inundation zones before floodwater arrives.

Field of the Invention

This invention relates to urban flood prediction and emergency routing, specifically to the combination of distributed IoT water level sensing within municipal stormwater infrastructure, physics-informed graph neural network modeling of pipe network hydraulics, and LiDAR-based surface flow simulation for real-time flood depth forecasting and dynamic route planning.

Background

Urban flash flooding is the deadliest weather hazard in the United States, responsible for an average of 88 fatalities per year (NWS 30-year average, 1995–2024), with over half of those deaths occurring in vehicles when drivers enter flooded roadways. The National Weather Service estimates that six inches of moving water can knock a person off their feet, and two feet of moving water can float a 3,000-pound vehicle. Global insured losses from urban flooding exceeded $35 billion in 2023 (Swiss Re sigma), with the share attributable to pluvial (rainfall-driven, non-riverine) urban flooding growing at approximately 7% annually as impervious surface area expands and rainfall intensity increases under climate change.

The fundamental problem is a temporal mismatch: current flood warning systems operate at basin scale (hours to days lead time, county-level spatial resolution) while urban flash floods develop at street scale (minutes to an hour, block-level spatial resolution). The National Weather Service issues Flash Flood Warnings based on Flash Flood Guidance (FFG) thresholds, typically at 1–3 hour averaging windows and county subdivision spatial units. By the time a warning is issued, water is already rising in low-lying underpasses and intersections. Municipal stormwater systems are designed to handle specific return periods (typically 10-year storms in the US, per EPA MS4 permit requirements), but operators have no real-time visibility into how close to capacity their pipe networks are running during a storm event.

Current approaches to urban flood prediction include:

The gap in the art is a system that: (a) instruments the stormwater pipe network itself (not just surface flood-prone locations) to observe hydraulic state in real time, (b) uses a physics-constrained neural network that combines the speed of ML inference with the physical fidelity of Saint-Venant equations, (c) predicts overflow locations and surface flood depths 15–60 minutes before flooding occurs, and (d) closes the loop to navigation and alert systems to reroute people away from danger before water arrives.

Detailed Description

1. Distributed Stormwater Level Sensor Network

The sensing layer instruments the municipal storm drain network at strategic locations: major trunk line manholes, junction chambers where multiple pipes converge, known bottleneck points (undersized pipes, siphons, flow restrictions), and surface catch basin inlets in flood-prone low areas. Target sensor density: one sensor per 200–500 meters of trunk line, plus all junction chambers and critical catch basins, yielding approximately 150–400 sensors for a typical city of 100,000–300,000 population.

Each sensor node comprises: a Maxbotix MB7389 HRXL-MaxSonar-WRMT ultrasonic rangefinder (20 cm to 5 m range, ±1 cm accuracy, IP67, unit cost $80) or, for lower-cost deployments, a JSN-SR04T waterproof ultrasonic module ($3–8, ±3 cm accuracy) with a conformal-coated PCB; a microcontroller (ESP32-C3 or STM32WL, unit cost $2–4) for data acquisition and LoRa communication; a LoRa radio (SX1262, 915 MHz ISM, unit cost $4); a lithium thionyl chloride (LiSOCl2) primary battery (Tadiran TL-5930, 3.6V 19 Ah, $12) providing 5–8 years of operation at one measurement per 10 seconds; and a stainless steel mounting bracket with cable gland for manhole or catch basin lid installation. Total BOM cost per node: $25–95 depending on accuracy tier. Installation requires no specialized equipment: the sensor is clamped to the underside of the manhole frame or catch basin grate, pointing downward at the water surface. Installation time: 15–30 minutes per node by a two-person crew.

The ultrasonic sensor measures the distance from the mounting point (at known elevation, from the municipal GIS manhole rim survey) to the water surface. Subtracting this distance from the sensor-to-invert distance (known from pipe network records) yields water depth in the manhole or catch basin. Each node transmits a 24-byte payload every 10 seconds during active rainfall (detected by a rate-of-change threshold on the water level signal) and every 5 minutes during dry weather, via a LoRa mesh network to one or more gateway nodes (Raspberry Pi with cellular backhaul) positioned at 1–3 km intervals.

2. Data Fusion Layer

The central inference server fuses three data streams:

3. Physics-Constrained Graph Neural Network (PC-GNN)

The core inference engine represents the stormwater pipe network as a directed graph G = (V, E) where vertices V represent junctions (manholes, catch basins, outfalls) and edges E represent pipe segments connecting them. Each vertex has a feature vector including: current measured water depth (if instrumented, otherwise NaN with a mask flag), manhole rim elevation, invert elevation, maximum storage volume, upstream contributing subcatchment area and imperviousness, and current rainfall intensity over the subcatchment. Each edge has a feature vector including: pipe diameter, length, slope, Manning's roughness coefficient, and computed full-pipe capacity via the Manning equation Q = (1/n)·A·R^(2/3)·S^(1/2) where A is cross-sectional area, R is hydraulic radius, and S is slope.

The GNN architecture uses 6 message-passing layers with 128-dimensional hidden states, based on a modified GraphSAGE aggregator. The key innovation is the physics-constrained loss function, which augments the standard MSE prediction loss with three penalty terms:

  1. Mass conservation penalty: For each junction node, the net inflow (sum of upstream pipe flows plus direct rainfall runoff) minus outflow (sum of downstream pipe flows plus any overflow) must equal the rate of change of stored volume. The loss term penalizes violations: L_mass = λ₁ · Σᵢ (ΔVᵢ/Δt − Qᵢ_in + Qᵢ_out)², where λ₁ = 0.1.
  2. Momentum constraint penalty: Pipe flow predicted by the GNN is penalized for deviating from the Manning equation estimate under gradually-varied flow conditions: L_momentum = λ₂ · Σⱼ (Q_pred,j − Q_manning,j)² for edges where the flow regime is subcritical (Froude number < 1). For pressurized or supercritical flow conditions, the penalty is relaxed. λ₂ = 0.05.
  3. Non-negativity constraint: Water depths and pipe flows must be non-negative: L_nonneg = λ₃ · Σ max(0, −h_pred)² + λ₃ · Σ max(0, −Q_pred)². λ₃ = 1.0.

The total loss is: L = L_MSE + L_mass + L_momentum + L_nonneg. The model is trained on a combination of: (a) historical sensor data from the deployed network (minimum 3–6 months of storm events), and (b) synthetic training data generated by running EPA SWMM simulations on the calibrated pipe network model under 500–2,000 synthetic storm hyetographs sampled from the site's intensity-duration-frequency (IDF) curves. Transfer learning from pre-trained models on similar-topology networks in other cities accelerates deployment: a model pre-trained on 5 cities' networks requires only 50–100 observed storm events for fine-tuning to a new city, compared to 200–500 events for training from scratch.

Inference speed: The GNN processes a full network update (all nodes and edges) in 50–200 ms on a single NVIDIA T4 GPU or equivalent, enabling 15–60 minute forward predictions with 10-second time steps in under 5 seconds of wall-clock computation. This is 200–1,000x faster than equivalent SWMM simulation.

4. Surface Overflow and Flood Depth Prediction

When the GNN predicts that water level at any junction will exceed the manhole rim elevation (or catch basin grate elevation) within the forecast horizon, the system computes surface overflow. The overflow volume rate Q_overflow = max(0, Q_in − Q_pipe_capacity) at the surcharging junction is distributed across the surrounding street surface using a pre-computed LiDAR-derived flow routing grid.

The surface routing uses a 1 m horizontal resolution DEM derived from municipal LiDAR surveys (typically available from USGS 3DEP at 1 m resolution for most US urban areas, or from state/county LiDAR programs). The DEM is pre-processed to: fill depressions and remove bridges/overpasses using a conditioned hydrologic DEM workflow; compute D8 flow direction and flow accumulation grids; identify surface flow paths (streets, ditches, channels) using curvature analysis; and pre-compute storage-depth-area curves for each street segment and intersection using the cross-sectional profile extracted from the DEM.

During a storm event, overflow volumes from surcharging junctions are routed across the DEM surface using a fast kinematic wave approximation: for each time step, overflow volume is distributed downstream along flow paths proportional to slope, with street segments acting as storage reservoirs. The resulting output is a time series of predicted flood depth at every street segment in the affected area, updated every 10 seconds, with a forecast horizon of 15–60 minutes. Predicted depths are classified into hazard levels: advisory (5–15 cm, pedestrian risk), warning (15–30 cm, vehicle hydroplaning risk), and danger (>30 cm, vehicle floating risk).

5. Dynamic Routing and Alert System

The flood depth forecast map feeds a routing API (REST/gRPC) that integrates with navigation platforms and alert systems:

6. Continuous Model Calibration and Uncertainty Quantification

The system continuously improves through online learning. After each storm event, observed water levels at all sensor nodes are compared to the GNN's predictions. Bayesian updating adjusts the model's uncertainty estimates for each pipe segment and junction. Manning's roughness coefficients (which change over time due to sediment accumulation, root intrusion, and pipe degradation) are recalibrated monthly using an ensemble Kalman filter that treats roughness as a latent state variable. The system also detects infrastructure changes (blockages, pipe collapses, new construction) as persistent model-observation divergences that trigger maintenance alerts.

Uncertainty is quantified using Monte Carlo dropout during inference (20 forward passes with dropout rate 0.1), producing prediction intervals rather than point estimates. The routing API reports flood depth as a probability distribution, enabling risk-appropriate responses: a 50% probability of 15 cm flooding triggers an advisory, while a 20% probability of 45 cm flooding triggers a danger warning.

7. Figures Description

Claims

  1. A system for real-time urban flash flood depth prediction comprising: a distributed network of water level sensors deployed within a municipal stormwater drainage infrastructure including storm drain inlets, catch basins, and manholes; a central inference server that receives real-time water level observations from said sensors; a graph neural network that models the stormwater pipe network as a directed graph with junctions as nodes and pipe segments as edges; wherein said graph neural network is trained with a physics-constrained loss function incorporating Saint-Venant equation conservation of mass and momentum penalty terms; and a surface flow routing module that computes predicted flood depth at street-segment resolution when the graph neural network forecasts pipe capacity exceedance at any junction.
  2. The system of claim 1, wherein the physics-constrained loss function comprises: a mass conservation penalty term that penalizes predicted net inflow minus outflow deviating from the rate of change of stored volume at each junction node; a momentum constraint penalty term that penalizes predicted pipe flow deviating from Manning equation estimates under gradually-varied flow conditions; and a non-negativity constraint term enforcing non-negative water depths and pipe flows.
  3. The system of claim 1, wherein the surface flow routing module uses a pre-computed LiDAR-derived digital elevation model at 1 meter or finer horizontal resolution, with D8 flow direction and flow accumulation grids, to route overflow volumes from surcharging junctions across the urban surface and compute flood depth at each street segment.
  4. The system of claim 1, further comprising a dynamic routing API that provides real-time flood depth forecasts as a GeoJSON layer to navigation applications, enabling rerouting of vehicles and pedestrians around predicted flood zones before surface water reaches hazardous levels.
  5. The system of claim 1, wherein sensor nodes comprise an ultrasonic rangefinder measuring distance to the water surface inside a manhole or catch basin, a LoRa radio module for low-power mesh communication, and a primary lithium battery providing multi-year autonomous operation without external power.
  6. A method for predicting urban flash flood depth and dynamically rerouting traffic comprising: instrumenting a municipal stormwater pipe network with distributed water level sensors at junction chambers and critical catch basins; fusing real-time water level observations with weather radar rainfall data and static pipe network attributes from a municipal GIS database; processing the fused data through a physics-constrained graph neural network that represents the pipe network as a directed graph and produces water level forecasts at each junction for a 15–60 minute horizon; computing surface overflow volumes and routing them across a LiDAR-derived digital elevation model to generate a street-level flood depth forecast map; and transmitting the forecast map to navigation applications, alert systems, and traffic signal controllers to dynamically reroute pedestrians and vehicles away from predicted inundation zones.
  7. The method of claim 6, wherein the graph neural network is trained on a combination of historical sensor observations and synthetic training data generated by running hydrologic-hydraulic simulations on the calibrated pipe network model under synthetic storm hyetographs sampled from the site's intensity-duration-frequency curves.
  8. The method of claim 6, further comprising continuous online calibration of pipe roughness coefficients using an ensemble Kalman filter that treats Manning's roughness as a latent state variable, recalibrated after each observed storm event.
  9. The method of claim 6, further comprising uncertainty quantification using Monte Carlo dropout during inference, producing prediction intervals for flood depth at each street segment, and triggering hazard-level-appropriate alerts based on the probability distribution of predicted depth.
  10. The system of claim 1, wherein a transfer learning approach enables deployment to new cities by pre-training the graph neural network on multiple existing city networks and fine-tuning with 50–100 observed storm events from the target city, reducing the training data requirement compared to training from scratch.

Prior Art References

  1. NWS Flood Fatality Statistics — Average 88 flood deaths per year in the US (30-year average)
  2. NWS Turn Around Don't Drown — 6 inches of moving water knocks a person down; 2 feet floats a vehicle
  3. Swiss Re sigma 2024 — Global insured flood losses exceeding $35B in 2023
  4. NWS Flash Flood Guidance definitions — Operational flash flood warning methodology
  5. EPA MS4 Stormwater Permits — Municipal separate storm sewer system permit requirements
  6. EPA SWMM — Storm Water Management Model, the industry-standard hydraulic simulation tool
  7. Guo et al. (2021), Water Research — Deep learning surrogate for urban flood simulation
  8. Bentivoglio et al. (2023), Journal of Hydrology — Graph neural networks for flood dynamics on unstructured meshes
  9. Flood Insights (Norfolk, VA) — Fixed flood sensor deployment in US cities
  10. NOAA NEXRAD Level III data — Weather radar precipitation products
  11. NOAA HRRR — High-Resolution Rapid Refresh model for short-range QPF
  12. USGS 3DEP — National LiDAR-derived elevation data program
  13. Rossman (2017), J. Hydraulic Engineering — SWMM hydraulic engine description and validation
  14. Nearing et al. (2024), Nature — Global prediction of extreme floods using graph neural networks (Google DeepMind)