MANGO: MANGROVE GLOBAL OBSERVATIONS –A DATASET AND BENCHMARK

ICLR 2026 Conference Submission25319 Authors

Keywords: Earth Observation, Mangrove

Abstract: Mangroves buffer coasts and store large amounts of carbon, yet they are vulnerable to storms and require reliable monitoring at global scale. Thresholded spectral indices break across sensors, seasons, and atmospheres, which limits their usefulness beyond local settings. Recent segmentation models are more promising but are difficult to train at scale because single-date imagery and labels are rarely paired and because models seldom exploit location context. First, we collect a globally distributed dataset, MANGO, that pairs one Sentinel-2 acquisition with each region–year label through a principled selection that balances agreement with the label and scene quality, and we provide country-disjoint splits together with co-registered geospatial embeddings. Second, we introduce a simple way to turn a global geospatial embedding into a small set of context channels that augment the optical bands and condition any backbone without architectural changes. Across strong convolutional and transformer baselines, this combination yields consistent gains on held-out countries and visibly cleaner maps, with sharper shorelines, better retention of small stands, and fewer false positives over turbid water, while adding minimal computational overhead. We release the dataset, the selection protocol, and the conditioning module to support reliable and scalable monitoring of coastal ecosystems.

Primary Area: datasets and benchmarks

Submission Number: 25319