Implicit Assimilation of Sparse In Situ Data for Dense & Global Storm Surge Forecasting

TL;DR

The paper tackles the challenge of short-term storm surge forecasting at ungauged coastal sites by proposing an implicit data assimilation framework that densifies sparse tide gauge observations with coarse ocean state reanalysis. It curates a global, multi-decadal dataset combining ERA5, GTSM, and GESLA-3 measurements to train and validate dense surge predictions, extending to unseen gauges. The authors evaluate conventional baselines, LSTM-based, and attention-based architectures within a densification scheme that includes auxiliary supervision of the coarse ocean state and in situ dropout, plus lead-time conditioning via FiLM. The resulting FiLM U-TAE model delivers the best hyperlocal performance (~16 cm MAE) and substantial gains in densification over GTSM, demonstrating feasibility of global, ungauged surge forecasting and setting the stage for operational deployment and satellite-augmented extensions.

Abstract

Hurricanes and coastal floods are among the most disastrous natural hazards. Both are intimately related to storm surges, as their causes and effects, respectively. However, the short-term forecasting of storm surges has proven challenging, especially when targeting previously unseen locations or sites without tidal gauges. Furthermore, recent work improved short and medium-term weather forecasting but the handling of raw unassimilated data remains non-trivial. In this paper, we tackle both challenges and demonstrate that neural networks can implicitly assimilate sparse in situ tide gauge data with coarse ocean state reanalysis in order to forecast storm surges. We curate a global dataset to learn and validate the dense prediction of storm surges, building on preceding efforts. Other than prior work limited to known gauges, our approach extends to ungauged sites, paving the way for global storm surge forecasting.

Figures (6)

• Figure 1: Overview: Our approach provides densified high-resolution storm surge forecasts (top) by implicitly assimilating inputs of sparse in situ tide gauge time series (top) with paired sequences of ocean (center) and weather state (bottom) re-analysis products. For additional supervision, coarse ocean state reanalysis maps (bottom) at coinciding lead time are also predicted.
• Figure 2: Data. Green & orange dots denote storm surge time series locations with records in 1979-2019, as pre-processed from the GESLA-3 collection of tide gauges haigh2023gesla. Dark lines indicate hurricane tracks in 2014-2019 as indexed by IBTRaCS knapp2010international. Pink markers highlight test split gauges, biased to points of landfall. Visualizations of the ERA5 grid and the irregular GTSM grid are omitted for brevity.
• Figure 3: Example data, one local sample per row. Input grouped at the left, targets at the right. Inputs: Time series of target (blue) and context gauges (grey), with target history only given in the hyperlocal setting. gauge locations and surge values. ERA5 pressure at mean sea level, wind at 10 m in u and v directions. Coarse GTSM input. Targets: Surge at target time, GTSM at target time.
• Figure 4: Experimental setup. Design of the densification and hyperlocal evaluation schemes, conceptualizing their respective inputs and outputs. The hyperlocal protocol focuses on forecasting of novel dynamics encountered at inference time, predicting surge at holdout target (green) and context gauges (blue) $L$ hours ahead. The generalization setup quantifies the goodness of models to broadcast predictions to ungauged locations, i.e. unknown gauges not contained in the input and solely used for evaluation.
• Figure 5: Location of holdout gauges impacts prediction performance. Average absolute errors in meters are color-coded and binned according to each sites' longitude and latitude coordinates, with gauge counts overlayed for each spatial dimension. Particularly challenging regions are the Gulf of Mexico, the Caribbean Sea and the Indian Ocean due to their extreme climate and resulting outsized surge dynamics.