Assessing Inundation Semantic Segmentation Models Trained on High- versus Low-Resolution Labels using FloodPlanet, a Manually Labeled Multi-Sourced High-Resolution Flood Dataset

Flooding impacts more people than any other environmental hazard, causing extensive economic and social impact. Leveraging satellite data and deep learning substantially improves flood monitoring and, potentially, management. However, deep learning efforts are frequently constrained by the limited availability of high-quality training and validation datasets, as well as the resolution, spatial coverage, and temporal limitations of inundation observation from public sensors. To address these challenges and contribute to the future development of geo-foundation models requiring extensive data and validation, we curate and publicly release FloodPlanet, a manually labeled inundation dataset [ 1 ]. FloodPlanet stands out for its manual annotations based on 3-m high-resolution commercial data from PlanetScope, and diverse ecoregions and flood event coverage. The dataset includes 366 hand-annotated labels, each 1,024 × 1,024 pixels, with corresponding Sentinel-1 and Sentinel-2 imagery, covering 19 global flood events from 2017 to 2020. Employing a “leave-one-region-out” cross-validation approach with a baseline UNet model, we achieved a mean intersection over union score (IoU) of 0.691 (SD: 0.227) for inundation detection across all events with PlanetScope, which is around 20% higher compared to Sentinel-1 and Sentinel-2 from the same event. Comparative analysis using PlanetScope labels to train models with Sentinel-1 and Sentinel-2 data revealed that FloodPlanet labels improve public sensor-based inundation detection by up to 15.6% (SD: 0.242) in IoU. These results imply that even if commercial data are too costly for near real-time inference applications, using some commercial data to train public sensor models could be an important lower-cost investment to increase accuracy.