🤖 machine learning

DELUGE: Towards Continental-Scale Daily Pluvial Flood Damage Prediction via Interpretable Conditioning on Foundation Model Embeddings

The paper introduces DELUGE, a multimodal deep learning framework that leverages interpretable conditioning on foundation model embeddings to achieve state-of-the-art, daily continental-scale pluvial flood damage predictions at ~1 km resolution, significantly outperforming traditional machine learning baselines on high-cost claims.

Original authors: Yuya Kawakami, Daniel Cayan, Dongyu Liu, Kwan-Liu Ma, Tom Corringham

Published 2026-07-20

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuya Kawakami, Daniel Cayan, Dongyu Liu, Kwan-Liu Ma, Tom Corringham

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Weather Detective's New Superpower

Imagine trying to predict exactly where a puddle will turn into a flood after a heavy rainstorm. This is the world of pluvial flooding, a type of disaster caused not by overflowing rivers or rising seas, but by rain that falls faster than the ground and gutters can drink it up. Unlike river floods, which have clear paths and gauges to track them, rain-driven floods are chaotic, local, and notoriously hard to pin down. For decades, scientists have tried to model these events using complex physics equations, like trying to simulate every drop of water in a bucket, but these methods are often too slow to run every day across a whole continent.

Enter Machine Learning, the branch of science where computers learn patterns from data instead of following rigid rulebooks. In recent years, a new tool called Foundation Models has emerged. Think of these as giant, pre-trained brains that have already "read" the entire world through satellite images and maps, learning what a city, a forest, or a swamp looks like without being explicitly told the rules. The big question researchers are asking is: Can we combine these super-smart, pre-trained brains with real-world insurance data to predict exactly where rain will cause damage, and can we do it in a way that humans can actually understand why the computer made that guess? This is crucial because if a computer says "flood here," emergency managers and insurance companies need to know if it's right before they send out alerts or pay out claims.

DELUGE: The Rain-Reading Robot with a Clear Conscience

Meet DELUGE (Daily Estimation of pLUvial flood damaGE), a new computer system designed to act like a hyper-observant weather detective. Its job is to look at the United States and predict, every single day, where rain will cause insured flood damage at a resolution of about 1 kilometer—roughly the size of a few city blocks. Instead of trying to cover the entire country equally, DELUGE focuses its energy on the top 100 areas that get hit the hardest by rain-related insurance claims, which accounts for about 81% of all the money spent on these floods.

The secret sauce of DELUGE isn't just that it's smart; it's that it's honest about how it thinks. Most advanced AI systems are "black boxes," meaning they give an answer but you can't see the gears turning inside. DELUGE, however, was built with "interpretability-by-design." Imagine a chef who doesn't just serve you a stew but also hands you a list of exactly how much salt and pepper they added to each specific pot. DELUGE does something similar for water. It uses two special "modulators" (think of them as adjustable lenses) that look at the local terrain and the "brain" of a foundation model (called AlphaEarth) to decide:

Value Modulator: How much does a specific amount of rain or wet soil matter in this specific neighborhood? (Maybe 1 inch of rain is a disaster in a sandy area but nothing in a swamp).
Temporal Modulator: How long does the memory of the rain last? (Does the flood happen instantly, or does the ground stay soggy for days, waiting for the next drop?)

By making these "lenses" adjustable and visible, the researchers can actually inspect the numbers the AI learned and say, "Ah, this neighborhood reacts to rain like a sponge, while that one acts like a rock."

How It Learned to See the Storm

To teach DELUGE, the researchers didn't just feed it raw data; they cleaned it up first. They used a massive database of insurance claims from the National Flood Insurance Program (NFIP). However, these claims are messy. People often report a flood days after it happens, or the address is vague. DELUGE's creators built a "correction pipeline" to fix this. They looked at weather records to see when it actually rained heavily near a claim and adjusted the dates accordingly. They also used geometry to shrink the vague address areas down to smaller, more precise spots on the map.

Once the data was clean, they trained DELUGE to look at four different types of information for every patch of land:

The Weather: How much rain fell, how wet the soil was, and how much snow was melting.
The Terrain: How steep the land is, how close it is to a drain, and what kind of soil it has.
The Exposure: How many houses and buildings are there?
The Vulnerability: What kind of houses are they? (Are they mobile homes? Do they have basements?)

The system then fuses all this information together to guess if a flood claim will happen on a specific day.

The Results: Smarter Than the Old Ways

When the researchers tested DELUGE, they pitted it against the standard tools used in the industry: Random Forest, XGBoost, and LightGBM. These are powerful, tree-based computer programs that have been the go-to for years. The results were clear: DELUGE won.

It didn't just win by a little bit; it outperformed the best tree-based models by 9% to 30% on a special score that cares most about the expensive, high-cost claims. This is a big deal because insurance companies and emergency managers care most about the disasters that cost the most money. DELUGE was particularly good at spotting these "heavy hitters."

The researchers also checked if the system was just memorizing the data or actually learning the rules of physics. They looked at the "lenses" (the modulators) the AI created for different neighborhoods. They found that the AI had learned logical patterns:

Urban areas (like dense cities) reacted very quickly to rain, just like you'd expect from concrete and asphalt.
Water-facing areas (near lakes or coasts) seemed to be sensitive to soil moisture over a longer period, suggesting the ground was already saturated.
Suburban areas showed a mix of behaviors.

This proved that the AI wasn't just guessing; it had learned hydrological rules that matched what scientists already know about how water moves.

What It Can't Do (Yet)

While DELUGE is a major step forward, the authors are careful to point out its limits. The system is only as good as the data it's fed. Since it relies on insurance claims, it misses all the floods that weren't reported or where people didn't have insurance. It also struggles in very dense cities where the damage happens on a scale smaller than the model can see (like a single street corner flooding while the next one stays dry). Additionally, the system is currently trained on the US; it hasn't been tested on other continents where the buildings and weather might be totally different.

The Big Picture

In short, DELUGE shows that we can build AI systems that are not only accurate but also transparent. By using "lenses" that can be inspected, the researchers turned a black-box AI into a tool that explains its own reasoning. This suggests a future where we can use these powerful foundation models to predict disasters across the entire continent, helping us prepare for the rain before it even starts to fall. It's not a magic crystal ball, but it's a very smart, very honest weather detective.

Technical Summary: DELUGE

Problem Statement
Pluvial (rainfall-driven) flooding accounts for 45% of National Flood Insurance Program (NFIP) claims in the United States but remains significantly harder to predict than riverine or coastal flooding. Existing approaches face three primary limitations: (1) static FEMA flood maps provide long-term hazard zones rather than dynamic daily damage estimates; (2) high-fidelity hydrodynamic models are computationally intensive and designed for fixed scenarios, not daily continental-scale operations; and (3) machine learning approaches are typically restricted to regional domains or rely on coarse, aggregated features. Furthermore, pluvial damage signals are noisy, derived from scattered insurance claims with significant spatial (redacted to Census Tracts/ZIPs) and temporal (reporting lags) uncertainties. There is a lack of a lightweight, daily, ~1 km resolution prediction system for insured pluvial damage across the Conterminous United States (CONUS).

Methodology
The authors present DELUGE (Daily Estimation of pLUvial flood damaGE), a multimodal deep learning framework designed to predict daily pluvial flood damage at ~1 km resolution. The system is trained on spatially and temporally corrected NFIP claims from 2017–2022, focusing on the top 100 highest-claim 75 km grid cells, which account for ~81% of total pluvial claims.

Data Preprocessing:
- Temporal Correction: A precipitation-guided pipeline aligns reported claim dates with NOAA AORC precipitation records. If a reported date lacks significant precipitation (>10 mm hourly max), the date is shifted to the nearest day within a ±3 day window that meets the threshold.
- Spatial Correction: Claims are localized by intersecting redacted Census Tract, ZIP code, and 0.1° grid cell data to create a polygon footprint. Claims are rasterized onto a 64×64 grid (~1.17 km/pixel) with confidence scores based on polygon area to mitigate the dominance of large rural claims.
- Negative Sampling: The dataset includes spatial negatives (same day, different pixel in a claim-containing grid) and temporal negatives (different day, same grid with >10 mm precipitation but no claim).
Architecture:
DELUGE is a multimodal CNN that fuses four branches: Hydrometeorology (hazard), Terrain (hazard), Exposure, and Vulnerability.
- Inputs: The model ingests a ~8×8 km patch (7×7 pixels) centered on the prediction pixel. Inputs include hourly precipitation, soil moisture, and snow water equivalent (73-hour window); terrain descriptors (HAND, TWI, Curve Numbers, soil properties); NFIP policy counts; and structural attributes from the National Structure Inventory. Crucially, it incorporates 64-dimensional Google AlphaEarth Foundation Model (AEF) embeddings at 10 m resolution.
- Interpretable Conditioning (The Core Novelty): Instead of concatenating AEF embeddings into a black-box encoder, DELUGE routes them through two parametric modules within the hydrometeorology branch:
  1. Value Modulator: Learns a per-patch, per-channel monotonic piecewise-linear warp ( $\phi$ ) of raw hydrometeorological values. This adjusts the relevance of specific value ranges (e.g., soil moisture levels) based on local terrain and AEF features.
  2. Temporal Modulator: Learns a per-patch temporal response kernel using a mixture of two Gamma distributions. This convolves the hydrometeorological time series to capture lag and decay timescales (e.g., immediate runoff vs. antecedent soil saturation) specific to the location.
- Fusion: The modulated hydrometeorology features are fused with terrain, exposure, and vulnerability features via a Transformer-based cross-attention mechanism, followed by an MLP classifier head.
Training: The model uses a spatial block holdout (splitting at the 75 km grid cell level) to prevent spatial leakage. It is optimized using focal loss weighted by claim confidence and a smoothness penalty on the Value Modulator parameters.

Key Contributions

DELUGE System: The first end-to-end learned system for daily, ~1 km pluvial flood damage prediction across the highest-claim regions of CONUS, outperforming standard tree-based baselines (Random Forest, XGBoost, LightGBM).
Uncertainty Correction Pipeline: Novel temporal and spatial correction procedures for NFIP claims that refine claim timing and localization, enabling their use as a daily training signal.
Interpretable Conditioning Modules: The introduction of Value and Temporal Modulators that condition foundation model embeddings (AEF) and terrain descriptors. These modules produce directly inspectable hydrological parameters (e.g., response peak lag, decay timescale, value relevance), offering "interpretability-by-design" rather than post-hoc attribution.
Physical Fidelity Validation: Demonstration via geography-blind clustering that the learned modulator parameters recover hydrologically coherent regimes (e.g., distinct urban vs. water-facing response patterns) consistent with physical expectations.

Results

Performance: Under a spatial block holdout, DELUGE achieves a PR-AUC of ~0.24, significantly outperforming the ~0.0025 no-skill baseline and tuned tree-based models by 9% to 30% on a dollar-weighted PR-AUC metric.
High-Cost Claims: The model excels at identifying high-cost claims. At a top-0.1% alert budget, DELUGE captures 62% of total claim dollars while identifying only 19% of claim events, outperforming baselines in this critical operational regime.
Ablation Studies: Removing hydrometeorological inputs causes a ~92% performance drop, confirming them as the dominant driver. Replacing the modulators with ConvLSTM or Transformer encoders results in comparable or slightly lower performance, validating that the modulator design achieves accuracy while adding interpretability.
Modulator Analysis: Clustering of learned parameters reveals distinct hydrological behaviors: Urban clusters show sharp reliance on recent precipitation, while water-facing clusters exhibit longer precipitation tails and wider soil moisture lookbacks, aligning with physical intuition.

Significance and Claims
The paper claims that DELUGE bridges the gap between computationally heavy physical models and coarse regional ML approaches, providing a scalable, daily prediction tool for insurers and emergency managers. The primary significance lies in the architectural novelty of the conditioning scheme: by routing foundation model embeddings through physics-shaped parametric modules, the authors demonstrate a transferable pattern for integrating geospatial foundation models into prediction tasks. This approach yields "interpretability-by-design," allowing stakeholders to inspect the hydrological response parameters (e.g., lag times, value thresholds) directly, rather than relying on opaque feature importance scores. The authors position this as a step toward physically meaningful, interpretable GeoAI, while acknowledging limitations regarding data noise (uninsured losses) and the need for higher-resolution inputs to resolve fine-scale urban damage.