Physically-based dimensionless features for pluvial flood mapping with machine learning

arXiv:2211.00636 · physics.geo-ph, stat.ML · Submitted 2026-08-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Physically-based dimensionless features for pluvial flood mapping with machine learning".

Jane: The paper was written by Mark S. Bartlett, Jared Van Blitterswyk, Martha Farella, Jinshu Li, Curtis Smith et al. from The Water Institute, Baton Rouge, LA, USA and Stantec, Ottawa, ON, Canada and Stantec, Flagstaff, AZ, USA and Stantec, Pasadena, CA, USA and Stantec, New York; NY; USA and Stantec; Lombard; IL; USA and Marquette University in Milwaukee; WI; USA.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: The authors of "Physically-based dimensionless features for pluvial flood mapping with machine learning" are doing something very clever by using the Buckingham Pi theorem to create these features, which is a complex idea but critical to understand. They are basically distilling the physics of water movement down into universal constants.

Jane: Think about it this way: they aren're not just plugging in raw data; they're defining the fundamental relationships between flow capacity, gravity, and landscape properties at a point. This lets them capture the pure physical laws governing pluvial flooding regardless of local environment or Meng’s specific input.

Lu: The use of the Buckingham Pi theorem is beautiful because it constraints the number of variables based on fundamental physics, rather than just letting an AI pick random features. It imposes scientific rigor onto what we usually treat as a black box ML process, which makes it much more interpretable for me as an AI researcher.

Meng: From my perspective, this mathematically forcing of the these physical constraints is exactly what I need to ensure the models are robust across geographical deployment scenarios and helps prevent unpredictable behavior when scaling up.

Lalam: It’s about creating a standardized language for flood risk, meaning we can move toward a more generalized approach that serves the broader public interest in flood resilience regardless of local geography.

Summary: Tom: The paper also gives us a very detailed summary of the methodology, and they aren're not just looking at the whole watershed at once; they are being very specific about the scale by looking at three distinct levels of detail.

Jane: They call these point, local, and non-local features, which is a very practical way to structure the data for our listeners. The point features capture things like immediate flow dynamics right there at a specific location.

Lu: The local features are crucial for capturing those small pathways like streets that can flood quickly even if the big river nearby isn't raging, and they represent subtle but important micro-level dynamics.

Meng: And non-local features handle the major arteries—the big streams—which is critical because of how much volume they can carry compared to those smaller channels, providing a comprehensive view of flow volume.

Lalam: This multi-scale approach ensures that we aren't missing any part of the flood event, which is essential for getting a complete picture of risk and preventing blind spots in our predictions.

Improvements: Tom: The paper makes a very bold claim that these dimensionless features perform significantly better than standard ML features, and that’s a huge claim to test. It suggests traditional models are too tied down to the specific environment they were trained in, right?

Jane: It suggests that our traditional models lack the ability to generalize well because they struggle with unforeseen conditions. The fact that we're testing them across different regions is a critical part of this finding, too.

Lu: But this method seems to allow the model to learn the *concept* of flooding rather than just learning patterns from the training data, which is a huge conceptual shift for AI researchers in us.

Meng: The cross-regional testing is what really stands out; seeing performance gains when training in Chicago and testing in New Jersey is a massive win for operational deployment, demonstrating true reliability.

Lalam: That improvement directly translates to faster, more reliable emergency response times when we are dealing with unknown areas or unexpected climates that the AI hasn's never seen.

Conclusion: Tom: We’ve seen how this paper uses the Buckingham Pi theorem to create these powerful features, and how they significantly boost ML model performance, which is a major achievement in flood science.

Jane: It’s a huge step toward making flood mapping a much more robust and generalized science than it currently is, moving beyond local solutions.

Lu: I think this is where the theoretical physics meets real-world application, creating something truly powerful for me as an AI researcher that has deep implications for modeling complex systems.

Meng: For my engineering perspective, this means we can deploy highly accurate warning systems that scale across different geographic areas without needing to redesign the entire model structure for every single location.

Lalam: This allows us to move towards a more equitable and efficient approach to hazard mitigation, ensuring that our technological advancements serve the broader public interest in flood resilience through all aspects of society.

Tom: So, as we wrap up this discussion on "Physically-based dimensionless features for pluvial flood mapping with machine learning," we're seeing that the future of reliable AI in hydrology is now looking much more generalized.

Jane: It’s a monumental achievement, ensuring that our models are not just tools for prediction but tools built upon fundamental physics.

Lu: I think the potential to generalize across diverse landscapes offers endless avenues for further exploration in AI applications.

Meng: The ability this will run in real-time is a practical game changer that needs to be fully utilized by engineers.

Lalam: It’s about building a future where risk assessment is more universal and efficient than ever before, guiding us toward better flood mitigation strategies for everyone.

Mark S. Bartlett, Jared Van Blitterswyk, Martha Farella, Jinshu Li, Curtis Smith, Anthony J. Parolari, Lalitha Krishnamoorthy, Assaad Mrad

The Water Institute, Baton Rouge, LA, USA · Stantec, Ottawa, ON, Canada · Stantec, Flagstaff, AZ, USA · Stantec, Pasadena, CA, USA · Stantec, New York; NY; USA · Stantec; Lombard; IL; USA · Marquette University in Milwaukee; WI; USA

physics.geo-ph, stat.ML

Submitted: 2026-08-21

Updated: 2026-08-25

Importance score: 81/100

The gist: " * Problem and Motivation: Rapid delineation of flash flood extents is critical for emergency resource mobilization.

Key concepts

Buckingham Pi Theorem
This theorem is used to distill the physics of water movement into universal constants. It forces fundamental physical relationships between flow capacity, gravity, and landscape properties. This imposes scientific rigor onto the model, ensuring it adheres to pure physical laws regardless of local environment.
Multi-scale Features
The methodology structures data at three distinct levels: point features capture immediate local dynamics; local features address small pathways like streets; non-local features manage major arteries. This ensures a comprehensive view of the entire flood event and prevents blind spots in predictions.
Generalization
This is the model's ability to perform reliably across diverse regions. By testing models trained in one area and applying them in another, researchers demonstrated true reliability. This allows for robust deployment of warning systems without needing to redesign the entire model structure.

Terminology

Summary

"

Problem and Motivation:

Rapid delineation of flash flood extents is critical for emergency resource mobilization. While Machine Learning (ML) approaches offer reduced computational demand compared to conventional high-resolution, 2D flood models, existing ML methods are limited by a lack of generalization to never-before-seen conditions. The study aims to improve ML model generalization by using a multi-scale framework based on dimensionless features that capture the similarity of the flooding process across regions.

Proposed Framework:

The core of the proposed solution is to utilize the Buckingham theorem to create a dimensionless formulation of pluvial flood mapping. This approach reformulates governing physical processes into a reduced-order, non-dimensional context, thereby inserting engineering and science metadata into the ML process.

Methodology (Dimensionless Feature Derivation):

The study applies the Buckingham theorem to two distinct physical processes:

  1. Storm Event Hydrology and Hydraulics (Flow Path Expansion): The governing equations are derived from the Saint–Venant equations, which describe conservation of mass (i.e., continuity) and momentum. These dynamics are analyzed at a point, where flooding occurs when the distance to water (z w) is less than or equal to zero. The resulting dimensionless groups (pi 1 through pi 4) are derived from the five independent variables: Q (flow rate), K v (conveyance velocity), z d (height above nearest drainage, for constraining the number of flooded points), A (cross-sectional area), and gS 0 (slope times gravity). The resulting dimensionless relationship is defined by:

pi 1 = f (pi 2, pi 3, pi 4

, the distance to sqrt the flood surface, z w,

  1. Long-term Hydrology (Saturated Areas): The analysis considers the assumptions of TOPModel regarding the dynamics of water table depth. The long-term recharge (q r) is balanced by outflow governed by Darcy’s Law. This leads to two dimensionless groups (1 and 2), which are related through the natural log function:

1 = (2)

Integration with Machine Learning:

The resulting dimensionless indices (pi and) are used as inputs to a logistic regression model. The overall process involves defining these features, estimating them, and training/testing the model. To capture the flood process in detail, these indices are calculated at three distinct scales: 1) at a point, 2) for smaller local flow paths (e.g., streets), and 3) for larger streams and rivers.

Implementation Details:

  • The ML process is applied to selected HUC 12 watersheds across the United States.

  • The study uses two distinct regions: Chicago (Case I training) and New Jersey (Case I testing). Case II involves training on both regions.

  • The model was trained using a logistic regression framework, comparing dimensional features against dimensionless features. Both sets of features were standardized.

Results and Performance:

The dimensionless feature ML model demonstrated superior performance compared to the traditional dimensional features:

  • Model Performance (AUC): The average AUC scores for dimensionless and dimensional features were consistently higher across all events (10-year, 100-year, and 1000-year). For Case I, the averages were 0.89 (dimensionless) vs. 0.84 (dimensional).

  • Model Performance (F beta-score): The F beta-score, which prioritizes minimizing false negatives (recall), also showed better performance with dimensionless features. In Case I, the average F2 scores were 0.65 (dimensionless) vs. 0.62 (dimensional).

  • Generalization: The most significant gains in performance occurred when the model was trained in one region and tested in another, confirming that dimensionless features provide an advantage compared to dimensional features with respect to generalization.

Discussion and Conclusion:

The study concludes that using the Buckingham theorem provides a path forward for developing more robust, scalable, and reproducible ML models. The dimensionless features improve generalizability by redefining the training space such that a larger number of test cases fall within the training set, reducing extrapolation. Furthermore, this approach is computationally efficient; once processed (2-3 hours for DEM resolution of 1- to 3-meters), the resulting data can be combined with runoff hyetographs and new ML results produced within minutes. This rapid mapping capability offers a significant improvement over current regional flash flood warnings, providing a path toward developing more robust, scalable, and reproducible ML models that can be applied to diverse and dynamic environments.

Improvements for AI systems

As a diligent AI researcher, I have analyzed this manuscript to identify foundational principles that can be applied to significantly enhance and generalize current AI systems in hydraulic and hydrologic modeling. The core insight—the use of physical invariants (dimensionless features) over purely empirical correlations—is a critical leap toward building robust, globally applicable models.

The following improvements leverage the insights of the Buckingham Theorem and multi-scale feature engineering to create a superior, generalized AI system.

Improvement: Replace standard, purely geomorphic features (e.g., Slope, DEM curvature) with Buckingham Invariants (i). These are dimensionless ratios derived directly from the governing physics (Saint-Venant and TOPModel equations).

  • Mechanism: The system captures the scaling similarity of a flood event—how water moves relative to its capacity, regardless of local units or absolute magnitude. Instead of learning that High Slope = Flooding, it learns that Gravitational Force / Frictional Resistance (a dimensionless ratio) = Flooding.

  • Goal: Eliminate domain-specific bias.

Improvement: Structure the input layer to explicitly encode flow dynamics across three defined scales: Point (pi p), Local (pi l), and Non-Local (pi nl).

  • Mechanism: The AI system simultaneously analyzes local, street-level drainage (low accumulation threshold) and regional, major riverine flooding (high accumulation threshold).

  • Goal: Provide the model with a comprehensive understanding of how small urban runoff interacts with large-scale fluvial processes, allowing it to predict localized flash floods and regional inundation in one pass.

Improvement: Adopt the F beta-score (beta=2) as the primary optimization metric within the training loop, rather than standard cross-entropy or AUC maximization.

  • Mechanism: The loss function is heavily weighted to penalize False Negatives (FN) twice as much as False Positives (FP). This forces the model to prioritize recall—ensuring that when a flood occurs, it is captured—over perfect precision.

  • Goal: Build a safety-critical system where missing a flood event (a false negative) is catastrophic, making the model inherently reliable for emergency response.

Improvement: Replace the simple Logistic Regression baseline with a Graph Convolutional Network (GCN) where the nodes are watershed points and edges represent flow connectivity defined by the derived features.

  • Mechanism: The GCN naturally processes spatial dependency. By feeding it physically-derived, dimensionless features (i), the the network learns not just if a point is flooded, but how the flood propagates through a network of connected catchment areas, respecting flow physics.

  • Goal: Achieve high spatial accuracy and capture complex flow dynamics far beyond what a simple tabular ML model can achieve.

The resulting system—a Physics-Informed, Multi-Scale Graph Network (PI-MS GNN)—will be capable of the following specific actions:

  1. Universal Transfer Learning: It will be able to deploy a single trained model (e.g, trained in Chicago) and achieve high predictive accuracy in entirely unrelated geographies (e.g., New Jersey), because the core relationship between flow dynamics (i) is invariant across the physical environment.

  2. High-Resolution Flash Flood Delineation: It can rapidly generate highly specific, pixel-level probability maps for pluvial flooding within minutes of receiving a weather forecast, providing actionable information far superior to generalized regional warnings.

  3. Prioritized Hazard Alerting: It guarantees maximum detection of actual flood events (high recall), ensuring that emergency resources are never diverted from areas that require immediate intervention due to the F beta-score optimization.

  4. Predictive Causal Interpretability: By analyzing which factors are driving a high probability prediction, the the system can provide an explainability layer—e.g., This area is flooding because its flow capacity ratio (pi 2) is low relative to its gravitational acceleration (gS 0), indicating insufficient drainage.

Related papers