R01 → Multi-Scale Feature Fusion for Predicting Water Disaster
Keywords:#CNN#Compond Water Disaster#Multi-Scale#Feature Fusion#
Year:2024 This work was published in the ACADIA 2024: Designing Change conference proceedings.
This study presents a multi-input, multi-scale feature fusion deep neural network for rapid prediction of compound water disaster risks during urban design. Using spatial representations of blue-green infrastructure and terrain elevation, the proposed model is trained on data derived from hydrodynamic simulations to generate high-accuracy water depth distribution maps at the block scale. Results show that multi-feature input significantly improves prediction performance, enhanced further by spatial-channel attention mechanisms and a feature pyramid structure. This method streamlines complex simulation workflows and serves as a fast, data-driven decision-support tool for evaluating flood resilience in urban design schemes.
Problem Statement
Conventional urban flood risk assessments rely on expert judgment and physics-based simulations, which are data-intensive, time-consuming, and difficult to integrate into early-stage design workflows. In particular, the absence of rapid, quantifiable tools hinders resilience evaluation of spatial layouts involving blue-green infrastructure during urban planning.
Preliminary Experiment
To assess the effect of input modality on prediction accuracy, preliminary experiments were conducted using single-input neural network models, trained separately with urban design images (RGB) and terrain elevation maps (grayscale). Results indicate that single-input models exhibit noticeable errors in simulating the spatial gradient and boundary of inundation areas, highlighting their limited capacity to capture flood patterns with reduced input complexity.
Core TECHNIQUES
The proposed model adopts an encoder-decoder framework with dual input branches that process terrain elevation maps (grayscale) and urban design images (RGB) in parallel.
Evaluation
Model performance was evaluated using SSIM (Structural Similarity Index), PSNR (Peak Signal-to-Noise Ratio), and average pixel error across RGB channels. The multi-input model achieved a mean SSIM of 0.91 and PSNR of 28.7 dB, outperforming the single-input baseline (SSIM: 0.82, PSNR: 24.1 dB).
Ablation experiments further validated the role of architectural components: removing both the SCAA and pyramid modules increased the average RGB channel error to 13.6, compared to 9.4 in the full model. These results confirm that the proposed attention and fusion mechanisms significantly improve the accuracy and fidelity of the predictions.
Application
This method enables rapid assessment of compound flood resilience during urban design by linking spatial layouts to predicted inundation maps through a trained neural network. It provides a data-driven foundation for evaluating flood risk in early-stage design scenarios.To explore the influence of blue-green spatial structures, four controlled experiments were conducted:
- Group 1: Adjusted river layouts while keeping green space constant.
- Group 2: Adjusted green space distribution with fixed river layout.
- Group 3: Simultaneously varied both river and green layouts.
- Group 4: Applied linear green corridor design to assess its effectiveness under specific hydrological structures.