Towards A Benchmark For Satellite Image Segmentation With Geospatial Foundation Models

Abstract

Geospatial foundation models, pre-trained on large amounts of satellite data, have showed promising performance in enhancing various downstream tasks. However, the benchmarking of such models has proved challenging due to the many fine-tuning settings that can be used. In particular, comparing foundation model performance on segmentation tasks can be challenging as there are many factors that are variable and can affect performance; such the decoder type, decoder depth, hyper-parameter settings, and more. This research investigates the performance of several foundations model architectures on geospatial segmentation tasks from the GeoBench benchmarking datasets. Experiments are conducted with different decoder and hyper-parameter settings for each foundation model. Finally, optimal settings for benchmarking foundation models on segmentation are prescribed.