Mangrove forests are some of the most vital ecosystems on the planet, protecting coastlines from natural hazards, supporting marine biodiversity, and filtering pollutants. They have long been threatened by land-use change, including conversion to aquaculture, but NASA satellites have seen them rebounding over the past decade, due to conservation efforts and the natural propagation of mangrove plants. A new study now shows how Earth observation (EO) foundation models, such as NASA and IBM’s Prithvi family of geospatial models, could support scalable monitoring of these dynamic coastal ecosystems.

Coastal Competition: Mangroves and Aquaculture

Researchers from Clark University in Massachusetts, NASA, and the University of Alabama in Huntsville who were involved in Prithvi development recently published a study in the Institute of Electrical and Electronics Engineers (IEEE) “Journal of Selected Topics in Applied Earth Observations and Remote Sensing.” The study evaluated how well Prithvi models map mangrove forests and aquaculture ponds under varying levels of labeled data availability and geographic and temporal domain shift.

Pond aquaculture, particularly shrimp farming, is a major driver of mangrove conversion in many regions around the world. Monitoring this change requires timely coastal maps, but producing the expert-labeled data needed to train supervised machine learning and deep learning models is costly and time-consuming.

To test whether Prithvi could reduce reliance on large labeled datasets, the researchers compared it to U-Net, a widely used deep learning architecture for remote sensing image segmentation. The research team assembled a training dataset from existing 2018 land use land cover (LULC) maps and pan-sharpened 15 m resolution Landsat-8 imagery covering ten South Asian countries and Ecuador. To test performance, the models were evaluated in two ways: first, on a reserved test set from the original regions to measure in-distribution performance, and second, on 2022 imagery from Mexico to evaluate how well they adapt to new environments.

Putting Geospatial AI to the Test

To measure each model’s mapping abilities, the team examined their success at semantic segmentation, which is the AI’s ability to accurately label pixels in the satellite images. The semantic segmentation task for this research involved assigning each image pixel to one of four classes: mangrove, aquaculture pond, water, or other land cover.

Field-verified environmental data is often scarce; therefore, the researchers designed a “stress test” to see how the AI models would perform under limited data availability scenarios. They trained the models using progressively smaller randomly selected subsets of the 6,540-image training dataset, scaling down from 100% to 50%, 25%, 12.5%, and just 5%, or 327 image chips (small, cropped sections of satellite images). By forcing both the Prithvi models and the baseline U-Net model to learn from these shrinking datasets, the researchers simulated real-world Earth observation workflows where labeled data is limited.

To isolate the value of pre-training, the researchers tested two approaches on each subset. They fine-tuned the Prithvi models from pre-trained weights and also trained them from scratch with randomly initialized parameters. U-Net was trained from scratch as the conventional baseline. All model configurations were then evaluated on the in-distribution and out-of-distribution test sets.

Where Foundation Models Have an Edge 

When trained with the full labeled dataset and tested within the same geographic domain, U-Net achieved the strongest overall segmentation performance. Its advantage narrowed, however, as labeled data became scarcer, and its performance was less stable when applied to imagery from Mexico. The 300-million-parameter model version of Prithvi maintained more consistent performance across different training data collections and domain types.