ZeroFlood: Flood Hazard Mapping from Single-Modality SAR Using Geo-Foundation Models

TL;DR

ZeroFlood introduces a Geo-Foundation Model-based framework for flood hazard mapping using SAR data, achieving high accuracy and demonstrating the potential of foundation models in data-scarce disaster prediction scenarios.

Contribution

The paper presents ZeroFlood, a novel framework leveraging Geo-Foundation Models for flood hazard mapping from single-modality SAR data, with a new dataset and improved performance techniques.

Findings

TerraMind achieves an F1-score of 88.36%.

GeoFMs outperform supervised baselines by over 3 percentage points.

Performance improves with the Thinking-in-Modality mechanism.

Abstract

Flood hazard mapping is essential for disaster prevention but remains challenging in data-scarce regions, where traditional hydrodynamic models require extensive geophysical inputs. This paper introduces \textit{ZeroFlood}, a framework that leverages Geo-Foundation Models (GeoFMs) to predict flood hazard maps using single-modality Earth Observation (EO) data, specifically SAR imagery. We construct a dataset that pairs EO data with flood hazard simulations across the European continent. Using this dataset, we evaluate several recent GeoFMs for the flood hazard segmentation task. Experimental results show that the best-performing model, TerraMind, achieves an F1-score of 88.36\%, outperforming supervised learning baselines by more than 3 percentage points. We shows the performance can be further improved by applying the Thinking-in-Modality (TiM) mechanism. These results demonstrate the potential of Geo-Foundation Models for data-driven flood hazard mapping using limited observational inputs. The dataset and experiment code are publicly available at https://github.com/khyeongkyun/zeroflood.