Utilizing a Geospatial Foundation Model for Coastline Delineation in Small Sandy Islands

TL;DR

The paper addresses coastline delineation on small sandy islands from low-resolution multispectral imagery, a task where existing tools often underperform. It evaluates NASA–IBM's Prithvi-EO-2.0 geospatial foundation model (300M and 600M) for image segmentation, fine-tuning on increasingly small labeled datasets derived from 225 Sentinel-2 images of two Maldives islands. The study finds that high shoreline delineation accuracy (F1 up to ~0.991) is achievable even with as few as 5–125 labeled images, with marginal gains from larger models and faster stability for the 600M variant. This demonstrates strong transfer learning capabilities of Prithvi-EO-2.0 for data-poor coastal monitoring and suggests practical deployment for historical shoreline analyses using low-cost, publicly available imagery.

Abstract

We present an initial evaluation of NASA and IBM's Prithvi-EO-2.0 geospatial foundation model on shoreline delineation of small sandy islands using satellite images. We curated and labeled a dataset of 225 multispectral images of two Maldivian islands, which we publicly release, and fine-tuned both the 300M and 600M parameter versions of Prithvi on training subsets ranging from 5 to 181 images. Our experiments show that even with as few as 5 training images, the models achieve high performance (F1 of 0.94, IoU of 0.79). Our results demonstrate the strong transfer learning capability of Prithvi, underscoring the potential of such models to support coastal monitoring in data-poor regions.

Figures (6)

• Figure A1: This is a run where we used CoastSat, a well-known coastline delineation toolkit, end to end to extract coastlines from an island in the Maldives. It can clearly be seen that the boundaries are varying extremely with tons of artifacts, even though the island shoreline itself is not changing so dramatically. Thus, these shorelines are not usable, highlighting the gap that our implementation fills when it comes to shoreline extraction.
• Figure A2: These are sample images of the two islands that our dataset is comprised of. The images were GeoTIFFS taken from Sentinel-2 and retrieved using Google Earth Engine. The images contain 6 optical bands -- R, G, B, Near Infrared (NIR), and Shortwave Infrared (SWIR) 1 and 2.
• Figure A3: Plotted Results of Prithvi 300M (Red) and Prithvi 600M (Blue), showing the relationships between performance and training dataset size. The x-axis has each of our sub training dataset sizes, while the y-axis has the range of score values, with the left plot showing F1 scores and the right plot showing IoU scores.
• Figure A4: These graphs show the performance of Prithvi 300M and Prithvi 600M, specifically the IoU scores, over the epochs. We can see performance levels out fairly quickly, around 10-20 epochs, with Prithvi 600M stabilizing more quickly.
• Figure A5: Figure A4 and Figure A5 are sample outputs of the fine-tuned versions of Prithvi. This Figure A4 shows a sample output of Prithvi 300M. Below, Figure A5 shows a sample output of the 600M version. Subtle differences can be seen between both of their predicted masks.