Flood Mapping from RGB imagery using a Vision Foundation Model

A new study examines whether a satellite-pretrained Earth observation foundation model can be adapted for flood mapping using airborne RGB imagery, a method that could support faster emergency response and damage assessment [1]. The research, posted to the arXiv preprint server on June 23, 2026, by Vladyslav Polushko, focuses on a model called Prithvi-2.0-UPN [1]. It combines the Prithvi-EO-2.0-600M Vision Transformer with a UPerNet decoder for binary water segmentation [2]. The work targets a persistent challenge: deep learning models for water segmentation often require large amounts of data to adapt when the scenery changes, such as during a new flooding event [2]. Airborne RGB imagery is attractive for this task because it can be collected rapidly and at low cost [2]. However, the foundation models now available for Earth observation were pretrained on satellite data, whose spatial resolution, viewing geometry, and radiometry differ from the nadir perspective of aerial photos [2]. The study investigates how to bridge that gap. The experiments used two RGB datasets: BlessemFlood21 and NeuenahrFlood [2]. When trained directly on these datasets, Prithvi-2.0-UPN reached state-of-the-art results on both [2]. In a second test, the model was trained on BlessemFlood21 and then applied to NeuenahrFlood without further training—a zero-shot setting. It outperformed baseline models, though the authors note the performance still indicated room for improvement [2]. A third experiment showed that fine-tuning the model with small shares of NeuenahrFlood training data led to rapid gains. Prithvi-2.0-UPN improved the fastest and nearly matched the performance level achieved when fully trained on the target dataset, demonstrating transfer capabilities [2]. The submitted paper file is 1,420 KB [1]. Flood mapping tools like this feed into broader geographic information systems (GIS), which integrate hardware, software, and data to store, analyze, and visualize geographic information [3]. GIS underpins applications in emergency management, urban planning, and environmental science by using location as a key index to relate otherwise unconnected data [3]. The arXiv repository where the study appears hosts more than two million open-access e-prints and receives roughly 24,000 new submissions each month, serving as a primary distribution channel for research in computer science, physics, and related fields [7].