Reproduce “Onboard Satellite Image Classification for Earth Observation: A Comparative Study of ViT Models” 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

Thanh-Dung Le, Vu Nguyen Ha, Ti Ti Nguyen, Duc-Dung Tran, Hung Nguyen-Kha, Luis M. Garces-Socarras, Juan Carlos Merlano-Duncan, Symeon Chatzinotas

Researchers at the Interdisciplinary Centre for Security, Reliability, and Trust (SnT) at the University of Luxembourg conducted a comparative study of Vision Transformer (ViT) and CNN models for onboard satellite land-use classification. The study identifies EfficientViT-M2 as an optimal choice, offering near-top accuracy (98.76% on EuroSAT) while consuming significantly less power (29.04 W) and demonstrating superior robustness to data corruptions common in satellite operations.

Key insights

  • Fine-tuning pre-trained backbones consistently yields higher accuracy than models trained from scratch, but larger models do not always guarantee better performance, emphasizing the role of architectural design.
  • Compact ViTs, specifically EfficientViT-M2, can achieve near-top accuracy while being significantly more energy-efficient and having a smaller memory footprint than other strong ViT contenders or even some lightweight CNNs.
  • Robustness to common satellite data corruptions (noise, blur, transmission loss) is a crucial, often overlooked, performance metric for onboard deployment, with some models exhibiting much more graceful degradation than others.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.