Reproduce “Segment Anything for Satellite Imagery: A Strong Baseline and a Regional Dataset for Automatic Field Delineation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Carmelo Scribano, Elena Govi, Paolo Bertellini, Simone Parisi, Giorgia Franchini, Marko Bertogna
Researchers at the University of Modena and Reggio Emilia and ABACO Group SpA adapted the Segment Anything Model (SAM) for automated agricultural field delineation, establishing it as a new baseline by outperforming Mask R-CNN with a 27.0% mAP50 on the AI4Boundaries dataset. They introduced the ERAgriSeg dataset for the Emilia-Romagna region of Italy and demonstrated the adapted model's strong temporal generalization to unseen data from subsequent years.
Key insights
- Zero-shot SAM performs poorly on satellite imagery for agricultural field delineation, highlighting the necessity of domain-specific fine-tuning for effective application in Earth Observation.
- Parameter-Efficient Fine-Tuning (PEFT) with LoRA effectively adapts SAM to specialized EO tasks, achieving optimal performance when applied to both the image encoder and mask decoder.
- Models fine-tuned on data from one year demonstrate robust temporal generalization to unseen data from subsequent years, indicating practicality for real-world, long-term agricultural monitoring systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.