A contrastive learning foundation model based on perfectly aligned sample pairs for remote sensing images

Self-supervised learning (SSL) facilitates the pre-training of foundation models without reliance on costly labeled data. Among SSL methods, contrastive learning (CL) excels at extracting robust semantic representations, even in the presence of complex interference. However, despite the success of CL in general computer vision, the significant domain gap necessitates specific adaptations for remote sensing imagery. To this end, we present a novel self-supervised method called PerA, which produces all-purpose remote sensing features through semantically perfectly aligned sample pairs. Specifically, PerA obtains features from sampled views by applying spatially disjoint masks to augmented images rather than random cropping. Our framework learns high-quality representations by ensuring consistency between teacher and student networks and predicting learnable mask tokens. Compared to previous contrastive methods, our method demonstrates higher memory efficiency and supports larger batch sizes due to the processing of sparse inputs. Additionally, the proposed method exhibits remarkable adaptability to uncurated remote sensing data and mitigates the impact of the potential semantic inconsistency. We also collected an unlabeled pre-training dataset, which contains about 5 million remote sensing images. We conducted experiments on multiple downstream task datasets and achieved performance comparable to previous state-of-the-art methods with a limited model scale, validating the effectiveness of our approach. We hope this work will contribute to practical remote sensing image interpretation.