Research map / Time series / Causal and generative time series
Off to new Shores: A Dataset & Benchmark for (near-)coastal Flood Inundation Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Curates the Global Flood Forecasting (GFF) dataset of 298 (near-)coastal flood events with multi-scale atmospheric, terrain, and Sentinel-1 inputs and benchmarks five baselines on global and coastal flood-extent forecasting tracks.
Problem
Flood inundation extent forecasting lacks datasets and benchmarks; existing work focuses on streamflow or rapid mapping rather than predicting flood extent maps ahead of time, especially for coastal regions.
Approach
A global multi-modal, multi-temporal dataset pairing coarse reanalysis/hydrological forcings (ERA5, ERA5-Land, GloFAS, HydroATLAS) with fine 10 m pre-event Sentinel-1 and terrain inputs, with flood-map labels generated by ensembling rapid-mapping models and a two-level context/local baseline platform conditioned on lead time via FiLM.
Key results
- U-TAE achieves the best overall F1 of 0.77 ± 0.04 on GFF, with water-class F1 of 0.57 ± 0.07.
- Best baselines reach F1 around 0.77 for flood forecasting, with the flooded (water) class remaining much harder (F1-W ~0.55-0.57).
- Baseline performance is generally higher at coastal regions (U-TAE F1 0.80 coastal vs 0.76 near-coastal/inland).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use U-TAE rather than the other benchmarked baselines | GFF Track 1 forecasting · GFF Track 1 | F1: 0.76 -> 0.77 | Helped |
| Use LSTM U-Net rather than U-TAE | GFF Kuro Siwo subset forecasting · GFF Kuro Siwo subset | F1: 0.72 -> 0.73 | Helped |
| Add GloFAS input to U-TAE (ablation) | U-TAE · GFF Track 1 | F1: 0.77 -> 0.76; 0.72 -> 0.73 | Mixed |
| Evaluate forecasts on coastal rather than near-coastal and inland areas | Benchmark baselines · GFF coastal versus near-coastal and inland | F1: 0.80 -> 0.76 | Helped |
| Add GloFAS input to U-TAE (ablation) | U-TAE · GFF coastal | F1: 0.80 -> 0.78 | Hurt |
| Add GloFAS input to U-TAE (ablation) | U-TAE · GFF near-coastal and inland | F1: 0.76 -> 0.75 | Hurt |
Surprising findings
- Performance on hand-annotated Kuro Siwo labels is only slightly lower than on the derived labels, with LSTM U-Net best on that subset (F1 0.73).
- Logistic regression on mono-temporal local data reaches F1 0.66, showing a non-trivial baseline gap relative to deep models of about 0.1.
Limitations
- Meteorological forcings are reanalysis rather than forecasts, possibly overestimating operational performance.
- Labels are generated by rapid-mapping models with residual artifacts and do not account for permanent water dynamics or non-hazardous submersion.
- Flooded pixels (minority class) remain hard to forecast.
Experimental setup
Datasets: Global Flood Forecasting (GFF) Kuro Siwo ERA5 ERA5-Land GloFAS HydroATLAS Sentinel-1 GRD CopDEM30 HAND ESA WorldCover Dartmouth Flood Observatory Caravan
Benchmarks: GFF Track 1 (global flood map prediction) GFF Track 2 (coastal vs near-coastal/inland) Kuro Siwo hand-labeled subset
Metrics: F1 F1 background F1 water standard deviation across 5 folds
Baselines: U-TAE LSTM U-Net 3DConv U-Net MaxViT U-Net logistic regression
Builds on: Kuro Siwo U-TAE SNUNet Caravan
Closest papers by meaning
- Early Detection of Extreme Storm Tide Events Using Multimodal Data Processing (AAAI 2024 · similarity 0.87)
- CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes (ICML 2025 · similarity 0.84)
- Kuro Siwo: 33 billion $m^2$ under the water. A global multi-temporal satellite dataset for rapid flood mapping (NeurIPS 2024 · similarity 0.82)
- Flood Insights: Integrating Remote and Social Sensing Data for Flood Exposure, Damage, and Urgent Needs Mapping (AAAI 2024 · similarity 0.82)
- CyPortQA: Benchmarking Multimodal Large Language Models for Cyclone Preparedness in Port Operation (AAAI 2026 · similarity 0.81)
- FIDLAR: Forecast-Informed Deep Learning Architecture for Flood Mitigation (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.