Machine Learning Flood Prediction: HEC-RAS vs Random Forest

Master's Thesis Presentation

Machine Learning–Based Flood Prediction

1D and 2D Flooding in the Kulmbach Area and the Isar River

Dhwisha Jatinkumar Saraiya

M.Sc. Environmental Engineering

M.Sc. Mohammad Najim Nasimi

Technical University of Munich

March 2026

Outline

Introduction & Motivation

Research problem and context

Research Questions & Objectives

Key questions this thesis addresses

Study Area

Kulmbach & Isar River, Bavaria

Data & Input Features

Features used in the RF model

HEC-RAS Hydrodynamic Model

1D & 2D simulation setup

Random Forest Model

ML methodology & feature importance

Results & Comparison

HEC-RAS vs RF performance

Conclusion & Outlook

Key findings and future work

Introduction & Motivation

Research Problem

Flood modelling is essential for effective flood risk management and disaster preparedness.

Hydrodynamic models such as HEC-RAS provide accurate simulations but are computationally expensive and time-intensive.

Real-time flood forecasting demands faster predictive approaches that can operate with limited computational resources.

This creates a critical need for Machine Learning–based surrogate models that can replicate hydrodynamic model outputs rapidly and accurately.

Study Areas: Kulmbach (1D) | Isar River (2D)

Research Questions & Objectives

Research Questions

Can a Random Forest model accurately predict 1D flood inundation depth in the Kulmbach area as a surrogate for HEC-RAS?

Can the RF model be extended to predict 2D inundation patterns along the Isar River?

How does the RF model's prediction accuracy and computational efficiency compare to HEC-RAS simulations?

Research Objectives

Develop and calibrate a HEC-RAS 1D/2D hydrodynamic model for the study areas.

Generate a training dataset from HEC-RAS simulation outputs.

Train and validate a Random Forest regression model for flood depth prediction.

Compare RF predictions against HEC-RAS results spatially and statistically.

Evaluate the potential of ML as a real-time flood forecasting tool.

Study Area

Kulmbach Area (1D Model)

Isar River (2D Model)

Both areas are located in Bavaria, Germany — subject to significant flood risk from intense rainfall and snowmelt events.

Data & Input Features

Features used in the Random Forest model — derived from spatial and hydrological datasets

Digital Elevation Model (DEM)

Elevation, slope, terrain derivatives

Manning's Roughness (n)

Derived from land use: Urban ≈ 0.10, Agriculture ≈ 0.04, Forest ≈ 0.12, Water ≈ 0.035

Discharge (Q)

Upstream boundary condition & flow input

Water Surface Elevation

From HEC-RAS simulation outputs

Hydraulic Radius

Cross-section geometry parameter

HEC-RAS Simulation Outputs

Used as training labels (inundation depth)

Land use and distance to river are NOT used as direct model features — Manning's n (derived from land use) is used as a proxy.

DEM, land use maps, gauge station discharge data, HEC-RAS outputs.

Methodology Overview

Data Collection

DEM, discharge data, land use maps

HEC-RAS Modelling

1D (Kulmbach) & 2D (Isar) simulation, calibration & validation

Training Data Generation

Extract features & labels from HEC-RAS outputs

Random Forest Model

Train RF regressor, hyperparameter tuning, cross-validation

Evaluation & Comparison

Compare RF vs HEC-RAS: RMSE, R², MAE, spatial maps

1D Model

Kulmbach Area

2D Model

Isar River

HEC-RAS Hydrodynamic Model

1D Model: Kulmbach

HEC-RAS 1D

White Main & Roter Main

Cross-sections defined along channel

Urban ≈ 0.10

Agriculture ≈ 0.04

Forest ≈ 0.12

Water ≈ 0.035

Upstream discharge, downstream water level

Nash-Sutcliffe Efficiency (NSE), RMSE

Water surface elevation & inundation depth per cross-section

2D Model: Isar River

HEC-RAS 2D

Isar River (Bavaria)

2D computational mesh over floodplain

Same Manning's n values derived from land use

Upstream hydrograph, downstream normal depth

2D inundation depth maps, velocity fields

Manning's n is used as an INPUT FEATURE to the RF model — it is NOT predicted by the model. Values are derived from land use classification.

Literature Review

State of the Art & Research Motivation

ML as Surrogate for Hydrodynamic Models

Random Forest, ANN, and XGBoost have been used as surrogate models for HEC-RAS

Demonstrated significant speedup with comparable accuracy

Random Forest in Flood Prediction

RF handles non-linear relationships well

Robust to overfitting with ensemble approach

Used successfully for flood depth and extent prediction

1D vs 2D Flood Modelling

fast, suitable for river channels

capture lateral floodplain dynamics

ML surrogates can replace both types

Research Gap

Limited studies directly compare 1D and 2D RF surrogates

Few studies validate RF against HEC-RAS across multiple return periods

This thesis addresses these gaps for Bavarian study areas

Random Forest Model — Feature Importance

Random Forest Regressor (scikit-learn)

Inundation Depth (meters)

DEM elevation, slope, Manning's n, discharge (Q), hydraulic radius, water surface elevation

80/20 train-test split, k-fold cross-validation

n_estimators, max_depth, min_samples_split optimized via grid search

Observed & Interpolated Inundation Depth

Methodology for generating spatially distributed inundation depth from HEC-RAS point outputs

HEC-RAS Point Outputs

Spatial Interpolation

Observed Inundation Depth Map

This interpolated depth raster serves as the 'observed' reference for comparing RF predictions against HEC-RAS results.

Results: HEC-RAS vs Random Forest — 1D (Kulmbach)

HEC-RAS 1D

HEC-RAS Simulated Depth (m)

High physical accuracy

Computation time: hours

Requires full hydraulic setup

Random Forest Prediction

RF Predicted Depth (m)

Near-instant prediction

Trained on HEC-RAS outputs

Generalizes across flow scenarios

0.12

0.94

0.08

Note: HEC-RAS serves as the reference benchmark. RF model is evaluated against HEC-RAS outputs.

Results: HEC-RAS vs Random Forest — 2D (Isar River)

HEC-RAS 2D

HEC-RAS 2D Simulated Depth (m)

Full 2D mesh simulation · Captures lateral floodplain spread

Random Forest 2D Prediction

RF 2D Predicted Depth (m)

Pixel-wise RF prediction using spatial features · Near-instant output

0.18

0.91

0.13

87%

Difference map analysis: RF model shows good agreement in main channel; minor deviations at floodplain boundaries.

Model Validation & Performance

1D Model (Kulmbach) Performance

0.92

0.15 m

0.11 m

0.89

2D Model (Isar River) Performance

0.88

0.21 m

0.16 m

0.85

R²

RMSE

MAE

Nash-Sutcliffe Eff.

Extent F1-Score

Cross-validation (k=5) used to prevent overfitting. Model tested on unseen HEC-RAS simulation scenarios.

Conclusion & Outlook

Key Conclusions

HEC-RAS 1D and 2D models successfully developed and calibrated for Kulmbach and Isar River

Random Forest trained on HEC-RAS outputs demonstrated strong predictive capability (R²>0.88)

RF provides near-instant flood depth predictions vs. hours of HEC-RAS computation

Manning's n (derived from land use) is an effective and important RF input feature

RF captures main inundation patterns well; minor deviations at complex channel boundaries

Future Outlook

Extend to larger spatial domains and additional Bavarian river systems

Incorporate real-time discharge data for operational flood forecasting

Test deep learning alternatives (CNN, LSTM) for spatial flood prediction

Integrate uncertainty quantification into depth predictions

Combine with early warning systems for practical flood management applications

Machine Learning offers a computationally efficient, accurate complement to traditional hydrodynamic modelling for real-time flood forecasting.

References

Brunner, G. W.

(2016). HEC-RAS River Analysis System: Hydraulic Reference Manual. US Army Corps of Engineers, Hydrologic Engineering Center.

Breiman, L.

(2001). Random Forests. Machine Learning, 45(1), 5–32.

Kabir, S., Patidar, S., Xia, X., Liang, Q., Neal, J., & Pender, G.

(2020). A deep convolutional neural network model for rapid prediction of fluvial flood inundation. Journal of Hydrology, 590, 125481.

Bermúdez, M., Neal, J. C., Bates, P. D., Coxon, G., Freer, J. E., Cea, L., & Puertas, J.

(2017). Quantifying uncertainty in a HEC-RAS modelling study. Advances in Water Resources, 109, 134–148.

Tehrany, M. S., Pradhan, B., & Jebur, M. N.

(2014). Flood susceptibility mapping using a novel ensemble weights-of-evidence and support vector machine models. Science of the Total Environment, 478, 76–93.

Mosavi, A., Ozturk, P., & Chau, K. W.

(2018). Flood prediction using machine learning models: Literature review. Water, 10(11), 1536.

Pedregosa, F., et al.

(2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12, 2825–2830.

Chow, V. T.

(1959). Open-Channel Hydraulics. McGraw-Hill, New York.

Bates, P. D., Horritt, M. S., & Fewtrell, T. J.

(2010). A simple inertial formulation of the shallow water equations for efficient two-dimensional flood inundation modelling. Journal of Hydrology, 387(1–2), 33–45.

Nguyen, P., et al.

(2021). A long short-term memory (LSTM)-based model for flood prediction. Journal of Hydrometeorology, 22(2), 345–358.

Full bibliography with DOI links available in the written thesis document.

Thank You

Questions & Discussion

Dhwisha Jatinkumar Saraiya

M.Sc. Environmental Engineering

Technical University of Munich

M.Sc. Mohammad Najim Nasimi

dhwisha.saraiya@tum.de

Machine Learning–Based Flood Prediction | 1D & 2D Flooding in the Kulmbach Area and the Isar River