Towards Precise Flood Prediction via Hierachical Terrain Attention and Multi-Scale Rainfall Guidance

Feifei Wang, Yong Wang, Shaoqing Chen, Bing Li, Qidong Huang

Abstract

With the deterioration of the global climate, rain-induced flooding has become increasingly frequent and destructive. To mitigate its impact, recent research has leveraged convolutional neural networks (CNNs) to predict flood maps. However, these methods often rely on global constraints to reconstruct raw pixels, overlooking the critical spatial information inherent in terrain features and complex rainfall patterns. We present a novel framework for precise flood map prediction that incorporates Hierarchical Terrain Spatial Attention (HTA) to focus on spatially salient terrain features and Multi-Scale Rainfall Embedding (MRE) to integrate rainfall information across different generation stages. Furthermore, we introduce a rainfall regression loss for both the generator and discriminator to provide additional supervision. Extensive evaluations on real catchment datasets demonstrate that our method significantly outperforms state-of-the-art approaches across various rainfall conditions.

Introduction

Predicting floods accurately and rapidly is essential for disaster mitigation. Traditional hydrodynamic simulations, while physically grounded, are computationally intensive and often too slow for real-time emergency response. Deep learning models, particularly those based on Generative Adversarial Networks (GANs), offer a promising alternative by treating flood prediction as an image-to-image translation task.

Despite their speed, existing deep learning models face two primary challenges. First, they often suffer from spatial uniformity, treating all geographic features with equal weight and failing to identify critical areas like steep slopes or narrow valleys where water accumulates rapidly. Second, they struggle with poor rainfall integration, often treating rainfall as a simple global variable that does not account for how complex patterns interact with local topography. This paper addresses these gaps by introducing a feature-aware architecture designed for robust generalization.

Methodology

Our proposed framework is built upon a GAN architecture that translates multi-channel terrain maps into water-depth maps. The core innovations include:

Hierarchical Terrain Spatial Attention (HTA)

To prevent the model from treating all terrain features equally, we integrate HTA modules into the encoder-decoder shortcuts. HTA generates dynamic weight maps that highlight spatially salient features—such as steep valleys or geographic bottlenecks—that are prone to rapid changes in water depth. By visualizing these layers, we observe that the model's focus shifts from local features to broader global structures as the network depth increases.

Multi-Scale Rainfall Embedding (MRE)

Rather than using a single input vector for rainfall, we employ a dedicated rainfall mapping network. This network extracts features at multiple scales, which are then injected into various stages of the decoder. This ensures that rainfall data influences both the coarse global structure and the fine-grained details of the generated flood map.

Rainfall Regression Loss

To ensure the model understands the causal relationship between rainfall and flooding, we add a regression branch to both the Generator ($G$) and the Discriminator ($D$). These components are tasked with regressing the original rainfall pattern from the generated flood images. This dual-purpose design forces the discriminator to learn more robust features, which in turn pushes the generator to produce more physically plausible results.

Experimental Setup

The study utilizes a public dataset from ETH Zurich, focusing on two distinct catchments: Zurich and Portugal. The input data consists of high-resolution terrain images across six channels: Digital Elevation Model (DEM), slope, aspect (cosine and sine), and curvature.

Visualization of the target flood map and the various terrain input channels used in the study.

The terrain data was processed into 10,000 patches of $256 \times 256$ resolution. For rainfall, 18 different one-hour patterns (hyetographs) were used, with 12 patterns reserved for training and 6 for testing to evaluate the model's ability to generalize to unseen conditions.

Results and Discussion

The proposed method significantly outperformed baselines such as U-flood, Pix2Pix, and TransUNet across all key metrics:

  • Accuracy: On the Zurich dataset, our model achieved a Mean Absolute Error (MAE) of 46.5, a substantial improvement over the previous best of 64.2 (U-flood). In Portugal, the MAE was 34.2, compared to 68.2 for U-flood.
  • Correlation: The model achieved an $R^2$ score of 89.1% in Zurich and 85.7% in Portugal, demonstrating a high degree of correlation between predicted and actual water depths.
  • Efficiency: While traditional hydrodynamic simulations can take up to 5 hours per condition, our model provides near-instant predictions once trained. Training on a single RTX 3090Ti takes approximately 2 hours and 53 minutes.

Conclusion

By moving beyond simple pixel reconstruction and incorporating hierarchical attention and multi-scale embeddings, our framework achieves state-of-the-art precision in flood map prediction. Future work will focus on expanding the model to handle temporal dynamics, allowing for the prediction of the entire evolution of a flood event over time, and reducing the dependency on high-resolution geographic data in regions where it may be scarce.

Summary updated

Abstract

With the deterioration of climate, the phenomenon of rain-induced flooding has become frequent. To mitigate its impact, recent works adopt convolutional neural network or its variants to predict the floods. However, these methods directly force the model to reconstruct the raw pixels of flood images through a global constraint, overlooking the underlying information contained in terrain features and rainfall patterns. To address this, we present a novel framework for precise flood map prediction, which incorporates hierarchical terrain spatial attention to help the model focus on spatially-salient areas of terrain features and constructs multi-scale rainfall embedding to extensively integrate rainfall pattern information into generation. To better adapt the model in various rainfall conditions, we leverage a rainfall regression loss for both the generator and the discriminator as additional supervision. Extensive evaluations on real catchment datasets demonstrate the superior performance of our method, which greatly surpasses the previous arts under different rainfall conditions.
BibTeX
@inproceedings{wang2023towards,
  author = {Feifei Wang and Yong Wang and Shaoqing Chen and Bing Li and Qidong Huang},
  title = {{Towards Precise Flood Prediction via Hierachical Terrain Attention and Multi-Scale Rainfall Guidance}},
  booktitle = {ICIP},
  year = {2023},
  url = {https://openreview.net/forum?id=boEpkjZtaD},
}

Loading related content...

Figure