Technical Summary: Uncovering Insights of Compound Flooding with Data-Driven AI

Problem Statement

Compound flooding, driven by nonlinear interactions among multiple hydrometeorological factors (e.g., rainfall, tides, groundwater, and human water management), poses a significant challenge to hazard prevention. While existing forecasting approaches—both physics-based and data-driven—have advanced, they often suffer from critical limitations:

• Physics-based models (e.g., HEC-RAS) require extensive high-resolution data (terrain, canal networks) that are difficult to standardize across regions, limiting large-scale comparative analysis.
• Data-driven models (e.g., LSTMs, Transformers) frequently treat compound flooding as a univariate temporal analysis task. They often emphasize temporal patterns while underexploring how multiple interacting factors jointly shape flood dynamics, leading to oversimplified scientific insights.
• Data Gaps: Existing datasets often lack integrated observational frameworks that simultaneously capture surface water, groundwater, rainfall, and human-controlled hydraulic operations, making it difficult to unravel the relative importance of these interacting drivers.

Methodology

1. Dataset Construction: SF2Bench

The authors constructed SF2Bench, a comprehensive observational time series dataset for South Florida, a region characterized by porous geology and dense water management infrastructure.

• Scope: Covers 67,349 km² from 1985 to 2024.
• Sources: Aggregates data from 2,452 monitoring stations operated by the South Florida Water Management District (SFWMD).
• Variables: Integrates heterogeneous physical and anthropogenic variables:
  • Surface water stage (993 stations).
  • Groundwater levels (582 stations).
  • Rainfall (349 stations).
  • Human-controlled operations: Pump rotational speeds (99 stations) and gate opening levels (429 stations).
• Preprocessing: Raw data from DBHYDRO was aligned to an hourly grid. Continuous variables (water levels) used linear interpolation for missing values, while stochastic event data (rainfall, pump/gate operations) were filled with zeros to represent the absence of events.

2. Experimental Design

The study reframes water stage forecasting not merely as a prediction task but as a system identification problem to uncover underlying dependencies.

• Models Benchmarked: Six categories of time series architectures were evaluated:
  • MLP-based (NLinear, TSMixer).
  • CNN-based (ModernTCN, TimesNet).
  • RNN-based (DeepAR, DilatedRNN).
  • GNN-based (FourierGNN, StemGNN).
  • Transformer-based (PatchTST, iTransformer).
  • LLM-based (GPT4TS, AutoTimes).
• Evaluation Metrics:
  • Standard: Mean Absolute Error (MAE) and Mean Squared Error (MSE) for general accuracy.
  • Extreme Events: Symmetric Extremal Dependence Index (SEDI) to specifically evaluate performance on tail events (flood peaks).
• Ablation Studies: Systematic removal of input factors (groundwater, rainfall, human control) and variation of spatial context (radius of surrounding stations) and temporal context (look-back window length) to isolate driver contributions.

Key Findings

1. Limitations of Temporal-Only Models

Models relying solely on temporal dynamics (e.g., standard RNNs or Transformers treating variables as a sequence) struggle to represent critical multi-factor interactions during compound events. While some Transformer-based models (PatchTST, iTransformer) achieved the best overall MAE, simpler MLPs often performed competitively, suggesting that complex temporal dependencies are not the sole driver of accuracy.

2. Dominance of Subsurface Saturation

A critical physical insight emerged from factor ablation: Groundwater levels are the dominant predictor of flood severity in this porous coastal region.

• Removing groundwater data caused a significantly larger degradation in model performance (MSE and MAE) than removing rainfall or human control data.
• For models like iTransformer and TimesNet, providing groundwater as the sole auxiliary input yielded the highest performance gains, indicating it acts as a "long-term memory" component governing surface water dynamics.

3. Spatial vs. Temporal Context

• Spatial: Incorporating data from surrounding monitoring stations within a finite effective radius provides critical causal context. Performance gains saturate once the immediate hydrological catchment is covered.
• Temporal: Extending the temporal history (look-back window) generally offers more substantial accuracy improvements than spatial expansion, particularly for capturing cumulative rainfall effects. However, beyond a certain point (e.g., >1 day for iTransformer), diminishing returns or overfitting occur.

4. Extreme Event Prediction

There is a decoupling between average accuracy and extreme event prediction. Models like FourierGNN demonstrated strong SEDI scores (capturing flood peaks) despite weaker average MAE/MSE, highlighting that standard error metrics may obscure a model's ability to capture tail dynamics.

5. Case Study: Hurricane Ian

In a case study of Hurricane Ian (2022), transformer-based models showed improved responsiveness to rapid water level increases compared to RNNs and CNNs. However, some models exhibited overestimation or sensitivity to noise under highly non-stationary conditions, reinforcing that accurate compound flood forecasting requires robust temporal alignment and stability.

Significance and Claims

The paper claims to offer new insights into the mechanisms of compound flooding by using data-driven models as tools for scientific inquiry rather than just prediction engines.

• Paradigm Shift: The study challenges rain-centric and sequence-dominated forecasting paradigms, suggesting that compound flooding is governed more by spatially coupled system states (specifically subsurface saturation) than by long-term temporal dependencies alone.
• Scientific Utility: By framing the dataset and analysis around causal discovery, the work provides a blueprint for understanding multi-hazard drivers in coastal environments.
• Future Direction: The authors suggest that future early-warning systems should be designed to be more physically grounded, explicitly integrating spatial context and subsurface states, and that interpretability frameworks are needed to understand how models weigh competing drivers (e.g., storm surge vs. river discharge).

The authors explicitly state that their findings are demonstrated in South Florida but reveal fundamental insights applicable to understanding compound flood mechanisms in other coastal regions. They do not claim to have solved the prediction problem entirely but rather to have identified the specific data and modeling gaps that currently hinder reliable extreme event forecasting.