On the Effectiveness of Pretrained Models for Remote Sensing

This thesis studies when natural-image pretraining versus satellite pretraining is most effective for remote sensing. It evaluates eight ResNet and ViT models across fine-tuning on BigEarthNet, linear probing on EuroSAT, and sample efficiency experiments, plus feature space analysis. Natural-image pretraining produces more separable, high-dimensional representations that excel in linear probing and high-data settings, while satellite pretraining yields compact features that improve low-data performance. The results clarify how domain alignment and data availability shape transfer performance.

Recommended citation: Declan Kutscher (2025). On the Effectiveness of Pretrained Models for Remote Sensing. M.S. thesis, Department of Computer Science, University of Pittsburgh.