Off to new Shores: A Dataset & Benchmark for (near-)coastal Flood Inundation Forecasting

Brandon Victor, Mathilde Letard, Peter Naylor, Karim Douch, Nicolas Longépé, Zhen He, Patrick EbelNeurIPS 2024

TL;DR

This paper introduces the Global Flood Forecasting (GFF) dataset, the first large-scale, multi-modal benchmark designed specifically for predicting future flood inundation maps. By bridging the gap between coarse weather reanalysis and high-resolution satellite imagery, the authors provide a platform to move from "mapping what is happening" to "forecasting what will happen" in vulnerable coastal regions.

While the remote sensing community has become highly proficient at rapid mapping (using satellites to see where water is currently sitting) and streamflow forecasting (predicting how much water will flow through a river), there is a critical missing link: spatial flood forecasting.

Rapid Mapping vs. Forecasting

Currently, disaster relief agencies often lack high-resolution maps showing exactly which neighborhoods or coastal areas will be underwater before a storm hits. Most existing datasets focus on 1D river gauges or post-event detection. Furthermore, (near-)coastal regions—where population density is high and flood drivers like storm surges and heavy rain often compound—are historically underrepresented in global flood benchmarks.

A Two-Level Pipeline for Scale Discrepancies

The authors propose a novel two-level pipeline to handle the massive discrepancy in data scales. Weather data is often provided at a coarse ~5.5km resolution, while actionable flood maps require a much finer ~10m resolution.

  1. Context Network (Top): Processes a spatio-temporal sequence of coarse data (ERA5 weather, GloFAS hydrology) over a 20-day window to capture the "big picture" of the developing flood event.
  2. Local Network (Bottom): Takes the embedding from the context network and combines it with high-resolution static data (Digital Elevation Models, HAND maps) and pre-flood Sentinel-1 radar images.
  3. Conditioning: The model is conditioned on a specific "lead time" ($d$ days) using Feature-wise Linear Modulation (FiLM). This allows a single model to predict flood extents for different future dates.

The study benchmarks several state-of-the-art architectures, including U-TAE (Spatio-Temporal Attention), LSTM U-Net, and Vision Transformers (MaxViT).

The Global Flood Forecasting (GFF) Dataset

The core contribution of this work is the GFF dataset, built to train and evaluate these predictive models.

Global Distribution of GFF Dataset ROIs
  • Scope: Features 298 Regions of Interest (ROI) across 6 continents and 13 climate zones, covering major flood events from 2014 to 2020.
  • Modality: Combines six distinct sensors and products, including ERA5/ERA5-Land (weather), HydroATLAS (basin attributes), Sentinel-1 (radar), CopDEM30 (elevation), and GloFAS (river discharge).
  • Scale: Contains 163,873 tiles (224x224 pixels) at 10m resolution.
  • Targets: Provides high-quality flood inundation maps derived from an ensemble of rapid-mapping models and expert annotations.

Key Insights and Results

  • The "Scale Gap" is the Primary Challenge: The hardest part of flood forecasting isn't just modeling the physics; it's the engineering challenge of fusing 5.5km-resolution atmospheric dynamics with 10m-resolution topography.
  • Coastal vs. Inland Predictability: Surprisingly, baseline models performed slightly better in coastal regions ($F1 \approx 0.80$) than inland ($F1 \approx 0.76$). The authors suggest this might be due to the predictable nature of coastal wetlands or the distinct, massive signatures of cyclone-driven events.
  • LSTMs Still Rule the Water: Mirroring successes in 1D streamflow modeling, LSTM-based U-Nets proved highly competitive, showing that the temporal "memory" of soil moisture and rainfall is vital for spatial forecasting.
  • Top Performer: The U-TAE (Spatio-Temporal Attention) model achieved the best overall performance with an F1 score of 0.77.
  • Class Imbalance: While models are excellent at predicting the "background" (non-flooded areas, $F1 \approx 0.97$), predicting the actual flooded pixels remains difficult ($F1 \approx 0.57$ for the best model), highlighting significant room for future innovation.

Limitations and Future Work

  • Reanalysis vs. Reforecasts: The dataset uses ERA5 reanalysis (which incorporates observations from the time of the event) rather than actual forecasts. In a real-world deployment, using actual weather forecasts might lead to lower performance due to inherent forecast errors.
  • Defining a "Flood": The definition of a flood can be subjective (e.g., are seasonal swamps considered "floods"?). To mitigate this, the dataset focuses specifically on hazardous events reported by the Dartmouth Flood Observatory.
  • Hydrological Complexity: While the model uses elevation and drainage maps (HAND), it doesn't explicitly model complex urban drainage systems or man-made flood defenses (like levees), which can drastically alter local outcomes.

Data and Code: The authors have released the GFF dataset and benchmarks at https://github.com/Multihuntr/GFF under a CC0 license.

Summary updated

BibTeX
@inproceedings{victor2024off,
  author = {Brandon Victor and Mathilde Letard and Peter Naylor and Karim Douch and Nicolas Longépé and Zhen He and Patrick Ebel},
  title = {{Off to new Shores: A Dataset \& Benchmark for (near-)coastal Flood Inundation Forecasting}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2024},
}
Figure