Research map / Time series / Causal and generative time series
Early Detection of Extreme Storm Tide Events Using Multimodal Data Processing
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.
Proposes a lightweight multimodal RNN+GNN model with periodic timestamp encoding that fuses sensor data, physics-based numerical forecasts (ROMS, SOFS) and satellite/reanalysis data to forecast sea-surface height and wave height for storm tides in Santos, Brazil.
Problem
Early detection and forecasting of extreme storm tide events in Santos, Brazil, from irregularly sampled, missing-data-prone multimodal sensor, numerical model, and satellite data, where high-resolution physics models are computationally expensive.
Approach
Combines independent per-series RNNs with time-gap positional encoding (sinusoidal, relative to the reference time) and a regularized heterogeneous GAT for information sharing, so irregular and missing series are handled without imputation and at low computational cost.
Key results
- Full model reaches IoA 0.96 for SSH and 0.65 for WH with all inputs (MAE 0.08 m SSH, 0.18 m WH) on 2022 test data.
- Removing time encoding lowers SSH IoA to 0.72 at 40% missing data versus 0.92 for the full model.
- Model has about 10.3M parameters, trains in 22.89 hours on a GeForce RTX 3060, and infers in 1.6 s on CPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add time encoding (ablation) | Full model · Santos storm tide forecast, 0% missing data | IoA (SSH): 0.94±0.10 -> 0.96±0.09 | Helped |
| add time encoding (ablation) | Full model · Santos storm tide forecast, 0% missing data | IoA (WH): 0.63±0.23 -> 0.65±0.23 | Helped |
| add RHGAT (ablation) | Full model · Santos storm tide forecast, 0% missing data | IoA (SSH): 0.92±0.12 -> 0.96±0.09 | Helped |
| add RHGAT (ablation) | Full model · Santos storm tide forecast, 0% missing data | IoA (WH): 0.52±0.25 -> 0.65±0.23 | Helped |
Surprising findings
- Adding only satellite-derived wave data (ERA5 waves) gives the largest WH gain, outperforming some numerical-model inputs.
Limitations
- Evaluated on a single location (Santos) and one test year (2022).
- Authors state the goal is not benchmarking; no comparison against standard baselines like GRU-D or RAINDROP is reported in experiments.
- Dataset includes a 3 km offset between simulation and measurement points, which is kept deliberately.
Experimental setup
Datasets: Santos Storm Tide Dataset ROMS simulations SOFS simulations Copernicus satellite altimetry (ADT, SSH) ERA5 waves and wind local SSH and wave/wind sensor measurements
Benchmarks: Santos Storm Tide Dataset (2022 test split)
Metrics: Index of Agreement (IoA) MAE
Baselines: ROMS SOFS GRU-D RAINDROP No time encoding variant No GNN variant
Builds on: Positional encoding (Vaswani et al. 2017) Heterogeneous graph attention network GRU-D RAINDROP
Closest papers by meaning
- Off to new Shores: A Dataset & Benchmark for (near-)coastal Flood Inundation Forecasting (NeurIPS 2024 · similarity 0.87)
- CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes (ICML 2025 · similarity 0.85)
- CyPortQA: Benchmarking Multimodal Large Language Models for Cyclone Preparedness in Port Operation (AAAI 2026 · similarity 0.83)
- FIDLAR: Forecast-Informed Deep Learning Architecture for Flood Mitigation (AAAI 2025 · similarity 0.83)
- Pluvial Flood Emulation with Hydraulics-informed Message Passing (ICML 2024 · similarity 0.82)
- RiverMamba: A State Space Model for Global River Discharge and Flood Forecasting (NeurIPS 2025 · similarity 0.82)
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.