ZeroFlood: A Geospatial Foundation Model for Data-Efficient Flood Susceptibility Mapping

TL;DR

This paper tackles flood susceptibility mapping in data-scarce regions where traditional physics-based models require dense inputs. It introduces ZeroFlood, a framework that fine-tunes Geospatial Foundation Models (GFMs) with Thinking-in-Modality (TiM) reasoning to predict flood susceptibility from unimodal Earth Observation data by leveraging paired EO and simulated flood maps from data-rich regions. A TerraMind/Prithvi-based evaluation shows that TiM improves robustness, with TerraMind-Large achieving a high F1 score around 67.21, demonstrating data-efficient, scalable FSM capability. The approach enables flood risk assessment in resource-limited settings and highlights the potential of foundation-model-based geospatial reasoning for disaster management.

Abstract

Flood susceptibility mapping (FSM) is vital for disaster prevention but remains challenging in data-scarce regions where hydrodynamic models require dense geophysical inputs. This work introduces ZeroFlood, a geospatial foundation model framework for data-efficient FSM. The approach fine-tunes Geospatial Foundation Models (GFMs) with Thinking-in-Modality (TiM) reasoning, enabling flood prediction from basic Earth observation data such as Sentinel-1 or Sentinel-2 imagery. Using paired EO and simulated flood maps from data-rich regions, ZeroFlood bridges data availability gaps through cross-modal representation learning. Experiments with TerraMind and Prithvi GFMs show that TiM enhances model robustness, with the TerraMind-Large configuration achieving an F1 score of 67.21. The results demonstrate the feasibility of foundation-model-based FSM as a scalable and data-efficient solution for flood risk management.

Figures (4)

• Figure 1: ZeroFlood framework for generating flood susceptibility maps in data-scarce regions using Geospatial Foundation Models (GFMs). Training data from data-rich areas are used to fine-tune large-scale GFMs. The Thinking-in-Modality (TiM) process is applied during fine-tuning and inference to compensate for missing modalities and enhance model performance.
• Figure 2: Illustration of the Thinking-in-Modality (TiM) process within a Geospatial Foundation Model. The process leverages information from additional modalities through pre-trained tokenizers trained on large-scale, multimodal geospatial datasets.
• Figure 3: Example of overlap between TerraMesh samples and flood simulation data. Flood simulation data (red pixels) and permanent water bodies (blue pixels) are shown along river networks. Rectangular grid boxes represent sample regions extracted from TerraMesh. Boxes meeting the data-quality criteria are shown in blue; otherwise, they are shown in red.
• Figure 4: Representative examples of flood susceptibility predictions using the TerraMind-L-TiM model. The Sentinel-1 RTC input and the predicted mask are shown in the middle, while the Sentinel-2 L2A reference image and ground-truth mask are displayed on the left and right, respectively.