Source-linked AI summary
Data-driven Flood Emulation: Speeding up Urban Flood Predictions by Deep Convolutional Neural Networks
Zifeng Guo, Joao P. Leitao, Nuno E. Simoes, Vahid Moosavi
TL;DR
Physically based simulations are too computationally demanding for large, high-resolution urban flood analyses. The paper uses a CNN to translate elevation, terrain, and rainfall information into maximum water-depth rasters, rather than repeatedly running simulations. The trained model takes 0.5% of cellular-automata model time and shows promising accuracy and generalization across rainfall variations.
Problem
Long computation times limit physically based flood simulations for large, high-resolution urban analyses and systematic scenario investigations.
Method
A CNN treats maximum water-depth prediction as image-to-image translation using terrain rasters and hyetograph inputs.
Results
0.5% of cellular-automata model time was required, with promising accuracy across shallow and deep waters and real rainfall events.
Takeaways & Limitations
The approach supports fast systematic flood investigations and analyses with generalization to rainfall variations not present in training data.
Takeaways & Limitations
The approach predicts maximum water depth for one specific catchment and does not model flow dynamics.
Abstract
from arXiv · showhide
Computational complexity has been the bottleneck of applying physically-based simulations on large urban areas with high spatial resolution for efficient and systematic flooding analyses and risk assessments. To address this issue of long computational time, this paper proposes that the prediction of maximum water depth rasters can be considered as an image-to-image translation problem where the results are generated from input elevation rasters using the information learned from data rather than by conducting simulations, which can significantly accelerate the prediction process. The proposed approach was implemented by a deep convolutional neural network trained on flood simulation data of 18 designed hyetographs on three selected catchments. Multiple tests with both designed and real rainfall events were performed and the results show that the flood predictions by neural network uses only 0.5 % of time comparing with physically-based approaches, with promising accuracy and ability of generalizations. The proposed neural network can also potentially be applied to different but relevant problems including flood predictions for urban layout planning.
1 Introduction
Urbanization and increasing rainfall intensity demand fast flood predictions, but physically based simulations remain too slow for large, high-resolution urban analyses. The paper addresses this bottleneck by using a CNN to generate maximum water-depth rasters from elevation data.
- Rapid flood predictions support systematic analysis of rainfall scenarios and could inform citizens about likely flood events.Rainfall spatial variation is also relevant when exploring different scenarios.
- Physically based simulations are computationally expensive, especially for large areas and high-resolution urban applications.Existing methods remain too slow for iterative analyses such as flood-driven optimization.
- The proposed method treats maximum water-depth prediction as image-to-image translation from elevation rasters using a convolutional neural network.The network learns from simulation data rather than running simulations for each prediction.
- The CNN was trained on simulation results from 18 designed hyetographs across three catchments and validated on designed and real rainfall events.
2 Previous Related Studies
Previous studies span physically based, simplified, cellular-automata, and data-driven approaches to flood and flow prediction. Their trade-offs motivate a CNN surrogate that can use raw elevation data for fast flood emulation.
- Flood simulation approaches: Physically based flood models solve differential equations and may represent flow in one or two dimensions.Reduced-complexity variants simplify momentum or flow-direction treatment.
- Flood simulation approaches: Non-physically based models provide approximate inundation predictions at lower computational cost when properties such as velocity are unnecessary.Cellular-automata flood models are a prominent example of this category.
- Data-driven methods: Artificial neural networks learn mappings from input data to outputs and have achieved substantial impact in complex computer-vision tasks.Computer-vision examples include image classification and object detection.
- Data-driven methods: Simulation-generated data have been used to train surrogate models that replace time-consuming flow simulations in interactive applications.
- Data-driven methods: Earlier flood-related neural and regression models used parameters from terrain or layout generators rather than raw elevation data.That input choice makes direct extension to other applications difficult.
3 Problem Statement
The paper focuses on predicting maximum water depth rather than flow dynamics because urban risk management and planning prioritize worst flooding cases. It initially targets one catchment while considering rainfall and terrain inputs.
- Urban flood applications emphasize worst-case water levels more than the dynamic process of flow.Accordingly, the study neglects flow dynamics and predicts maximum water depth.
- Maximum-water-depth prediction can address rainfall variation, spatial variation, or simultaneous terrain and rainfall variation.The paper identifies these as distinct application problems.
- The proposed approach predicts maximum water depth for one specific catchment from hyetographs and elevation data.
- The paper notes that flow velocity could potentially be included by adding extra output image channels.
4 Proposed Approach
The approach encodes terrain and rainfall information, trains a CNN on raster patches, and aggregates patch predictions into a catchment-wide water-depth raster. Its design combines multichannel terrain representation, a joint network, imbalance-aware loss, and overlapping-patch aggregation.
- Framework: Terrain and water-depth patches are randomly sampled for training, while an orthogonal grid supplies prediction locations after training.Rainfall hyetographs are sampled every 5 minutes over one hour as 12-dimensional vectors.
- Catchment representation: Terrain is represented with five channels—elevation, slope, aspect, curvature, and mask—rescaled and concatenated into a multichannel image.Adding these features made training significantly faster than using elevation alone.
- CNN prediction model: The joint model combines a convolutional autoencoder for terrain with a fully connected subnetwork for hyetograph data.The subnetworks combine at the latent layer before decoding water-depth values in metres.
- CNN prediction model: The network maps 256 × 256 × 5 inputs to 256 × 256 × 1 outputs using convolutional, pooling, upsampling, and fully connected layers.Small kernels preserve thin terrain structures while deeper layers extend the receptive field.
- Loss function: A weighted mean squared error addresses the imbalance caused by abundant no-water and shallow-water cells relative to deep-water cells.The weighting is based on simulated water depth, with c = −1 used in all tests.
- Patch aggregation: Patch predictions are aggregated over an orthogonal grid using mean, median, maximum, or no-overlap options.The default tests use grid size 128 and mean aggregation; overlap can reduce outlier effects but increases computation.
5 Experimental Setup
The experiments applied the framework to three catchments using simulated rainfall and evaluated computational time, accuracy, aggregation trade-offs, and generalization to real rainfall events.
- Study areas: The framework was applied to three catchments in Switzerland and Portugal using 1 m-resolution elevation rasters with GPU acceleration.Simulation, training, and validation were performed with GPU support.
- Simulation rainfall data: Each catchment used 18 one-hour design-rainfall simulations spanning return periods from 2 to 100 years, with one-third of each period held out for testing.The CNN was trained only on the remaining simulations and evaluated on the held-out set.
- Flood simulation models: The simulations used CADDIES for Zurich and Luzern and Infoworks ICM for Coimbra, retaining raw maximum-water-depth outputs without postprocessing.Users could apply reality-oriented thresholds, such as the 90th percentile, after prediction.
- Training data: Training and testing used terrain patches and water-depth patches sampled from each catchment, with 10,000 terrain patches and 180,000 water-depth patches per catchment.Separate models were trained for each catchment to examine generalization across hyetograph inputs.
- Evaluation and validation: Performance was evaluated by computational time, prediction accuracy, and generalization to hyetographs, including trade-offs among aggregation methods.Accuracy used mean absolute error and 2D histograms comparing predicted with simulated raster-cell depths.
- Evaluation and validation: Models were additionally validated on real rainfall events excluded from training after clipping events to 60 minutes and resampling them into 12-dimensional vectors.Results were reported with histograms and spatial plots.
6 Results
On the 100-year design rainfall event, the model substantially reduced prediction time while maintaining accurate water-depth predictions across catchments and water-depth ranges. It also generalized to real rainfall events, though localized errors occurred near sharp simulated depth changes.
- Prediction time: 0.5% of cellular-automata simulation time was required by the trained model for water-depth prediction across all three catchments.The no-patch-overlap option was fastest.
- Prediction time: Less than 6.2% separated mean- and maximum-value aggregation times on average, while median aggregation was slowest because it retained all data in memory.
- Patch aggregation: Different patch aggregation methods produced similar overall accuracy, although overlapping patches generally lowered MAE and median aggregation usually outperformed mean and maximum values.No overlap produced more under- and overpredictions, suggesting more outliers.
- Water-depth accuracy: The 2D histograms aligned closely with the diagonal from 0.0 m to 7.0 m, indicating accurate predictions in both shallow- and deep-water areas.The histograms used 0.1 m resolution and logarithmic colouring to show small shallow-water counts.
- Water-depth accuracy: Some simulated 0 m cells were overpredicted by up to 2–3 m, mainly where the model smoothed sharp water-depth changes associated with possible elevation-data artefacts.These high-error cells were considered small relative to the millions of cells in each catchment.
- Generalization: The model showed no significant shallow-versus-deep-water performance difference and accurately handled three real Coimbra rainfall events with one or multiple peaks.
7 Conclusions and Future Steps
The CNN approach substantially reduces flood-prediction time while retaining promising accuracy and generalization across catchments, water depths, and rainfall variations. Future work includes extending inputs and outputs and validating against observational data.
- 7 Conclusions and Future Steps: 0.5% of cellular automata-based model time was required by the trained CNN model.The approach was tested in three different catchment areas, with significant computational-time improvement.
- 7 Conclusions and Future Steps: CNN runtime is expected to increase linearly with the number of patches as raster resolution increases.This contrasts with physically based models, which could potentially require 10 times more computational time when raster resolution doubles.
- 7 Conclusions and Future Steps: High accuracy was achieved in both shallow- and deep-water areas despite imbalanced data.Tests with real rainfall events absent from training data suggested generalization to input rainfall variations.
- 7 Conclusions and Future Steps: Future work could use different input terrains, retain unchanged hyetographs, or add features such as water velocity.These combinations are identified as immediate next steps for the research.
- 7 Conclusions and Future Steps: Future studies could train and validate the model using observational data instead of simulations.Crowdsourcing and computer vision techniques are identified as making this direction increasingly possible.