Keywords: aerial imagery, foundation models, self-supervised learning, benchmark
TL;DR: A benchmark that measures generalization of remote sensing foundation models to unseen resolutions and bands
Abstract: Foundation models have significantly advanced machine learning applications across various modalities, including images. Recently numerous attempts have been made on developing foundation models specifically tailored for remote sensing applications, predominantly through masked image modeling techniques. This work explores the essential characteristics and performance expectations for a foundation model in aerial imagery. We introduce a benchmark designed to evaluate the model's performance as well as robustness to changes in scale and spectral bands of the input. Our benchmarks encompass tasks unique to aerial imagery, such as change detection and scene classification, and utilize publicly available datasets RESISC45, BigEarthNet, LEVIR-CD and OSCD. We evaluate recently proposed foundation models on the benchmark. Furthermore, we explore the impact of various design choices in pretraining and fine-tuning on the performance of the models on our benchmark. Specifically, we pretrain several variations of a self-distillation based self-supervised model on aerial imagery datasets, including one without scale-augmentations and another one with a pretrained mask decoder module.
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13991
Loading