Deep Neural Networks Measure Friction Factors from 3D Point Clouds to Improve Hydrodynamic Flood Models

Core Concepts

A laboratory-trained deep neural network can directly measure Manning's n from 3D point cloud data, improving the accuracy of 1D, 2D, and coupled 1D/2D hydrodynamic flood models compared to using land cover-derived values.

Abstract

The research presents a method to measure Manning's n, a key friction factor used in hydrodynamic flood modeling, directly from 3D point cloud data using a deep neural network (DNN) trained on laboratory experiments.
The key highlights are:

A 10-m non-recirculating flume was used to measure Manning's n on concrete slabs with varying surface roughnesses. 3D point clouds of the slabs were collected using a handheld lidar scanner and a flume-mounted bed profiling system.
A PointNet-based DNN architecture was adapted to predict Manning's n from the point cloud data, with data augmentation techniques used to expand the limited experimental dataset.
The trained DNN was applied to 3D point cloud data from a lidar survey of Houston, Texas to obtain high-resolution spatial measurements of Manning's n.
The lidar-derived Manning's n values were incorporated into 1D, 2D, and coupled 1D/2D hydrodynamic models of regulatory flood events and Hurricane Harvey in Houston.
Compared to using Manning's n values derived from land cover data, the lidar-based values showed improved agreement with validation data, including water depth measurements, flood extents from imagery, and flood insurance claims.
The impacts of friction factors were found to be significant for 1D and 2D models under riverine and pluvial flooding, while surge flooding was largely unaffected.
The results demonstrate the potential of using 3D point clouds and deep learning to provide a reliable, repeatable, and readily-accessible method to measure friction factors and improve the accuracy of hydrodynamic flood models.

Stats

The laboratory experiments measured Manning's n on three concrete slabs with varying surface roughnesses, with at least four different flow rates for each slab.
The mean Manning's n values for the three slabs were 0.025, 0.035, and 0.055.

Quotes

"Friction is one of the cruxes of hydrodynamic modeling; flood conditions are highly sensitive to the Friction Factors (FFs) used to calculate momentum losses."
"The lidar measurements result in better fit than the NLCD values for the Water Surface Elevation (WSE) timeseries at both gauges."
"Changing FFs significantly affected fluvial and pluvial flood models, while surge flooding was generally unaffected."

Key Insights Distilled From

Deep Neural Networks with 3D Point Clouds for Empirical Friction Measurements in Hydrodynamic Flood Models

by Francisco Ha... at arxiv.org 04-04-2024

https://arxiv.org/pdf/2404.02234.pdf

Deeper Inquiries

How can the presented DNN architecture be further optimized to improve its performance and reduce the observed scale dependence during inference?

The presented DNN architecture, based on PointNet, can be further optimized to enhance its performance and reduce scale dependence during inference. Here are some strategies for optimization:

Data Augmentation Techniques: Implement more advanced data augmentation techniques to increase the diversity and quantity of training data. This can include rotation, scaling, and translation of point clouds to expose the model to a wider range of variations.

Feature Engineering: Introduce additional features or transformations of the point cloud data that can provide more meaningful information to the DNN. This can help the model better capture the underlying patterns in the data.

Architecture Modifications: Experiment with different architectures or modifications to the existing PointNet structure. This can involve adding or removing layers, adjusting the number of neurons in each layer, or incorporating attention mechanisms to focus on important regions of the point cloud.

Regularization Techniques: Apply regularization techniques such as dropout or batch normalization to prevent overfitting and improve generalization of the model.

Hyperparameter Tuning: Fine-tune the hyperparameters of the DNN, including learning rate, batch size, and optimizer settings, to optimize the training process and improve performance.

Transfer Learning: Explore the possibility of using transfer learning by pre-training the DNN on a related task or dataset before fine-tuning it on the specific Manning's n prediction task. This can help leverage knowledge learned from other domains to improve performance.

How can the potential limitations of using point cloud data collected from different sensors, scanning patterns, and quality levels be addressed, and how can the DNN be made more robust to such heterogeneity?

The potential limitations of using point cloud data from different sources can be addressed by implementing the following strategies to make the DNN more robust to heterogeneity:

Normalization and Standardization: Normalize and standardize the point cloud data to ensure consistency in scale, orientation, and density across different sources. This can help mitigate the effects of variations in sensor types and scanning patterns.

Feature Alignment: Align features extracted from point clouds from different sensors or scanning patterns to a common reference frame. This can involve registering point clouds to a standard coordinate system to ensure compatibility.

Quality Assessment: Implement quality assessment techniques to evaluate the reliability and accuracy of point cloud data from different sources. This can involve identifying and filtering out noisy or erroneous data points to improve the overall quality of the input.

Ensemble Learning: Utilize ensemble learning techniques to combine predictions from multiple DNN models trained on point cloud data from different sources. This can help mitigate the impact of heterogeneity and improve the robustness of the predictions.

Adversarial Training: Incorporate adversarial training methods to train the DNN to be resilient to variations and perturbations in the input data. This can enhance the model's ability to generalize across different sensor types and scanning patterns.

Given the importance of friction factors in hydrodynamic modeling, how can the insights from this study on the differential impacts of friction on fluvial, pluvial, and surge flooding be leveraged to develop more advanced friction formulations in flood models?

The insights from this study on the impacts of friction factors on different types of flooding can be leveraged to develop more advanced friction formulations in flood models in the following ways:

Dynamic Friction Models: Develop dynamic friction models that can adapt to different flow conditions and flood types. Incorporate the differential impacts of friction on fluvial, pluvial, and surge flooding to create more nuanced and accurate friction formulations.

Multi-Parameter Friction Models: Introduce multi-parameter friction models that consider a combination of Manning's n, Strickler's K, and Chezy's C coefficients to capture the complex interactions between flow characteristics and surface roughness.

Machine Learning-Based Friction Estimation: Utilize machine learning techniques, such as DNNs, to predict friction factors based on point cloud data and other relevant features. This can provide a data-driven approach to friction estimation that accounts for the differential impacts observed in the study.

Integration of Remote Sensing Data: Integrate remote sensing data, such as high-resolution lidar point clouds, to improve the accuracy and spatial resolution of friction factor measurements. This can enhance the fidelity of friction formulations in flood models and enable more precise flood predictions.

Validation and Calibration: Validate and calibrate the advanced friction formulations using real-world flood data and observations. This iterative process can refine the models and ensure their reliability in simulating different flood scenarios accurately.

By incorporating these strategies and leveraging the insights from the differential impacts of friction on various types of flooding, more advanced and robust friction formulations can be developed to enhance the accuracy and effectiveness of hydrodynamic flood models.