Deep learning has shown remarkable success in analyzing grounded imagery, such as consumer photos due to large-scale human annotations are available for dataset, e.g., ImageNet. However, such extensive supervision is not the case for remote sensing data.
We propose a highly effective semi-supervised approach tailored specifically for remote sensing data. Our approach encompasses two key contributions. We adapt the framework from semi-supervised learning approach, such as FixMatch, to remote sensing data by designing a set of robust strong and weak augmentations suitable for this domain. By learning from actual labeled data, combining with pseudo-labeled data, yet address the pseudo-labeling imbalance, we leverage a recently proposed debiased learning approach to mitigate the bias in pseudo-labeling.
Validated by extensive experimentation, our simple semi-supervised framework with 30\% annotations delivers significant accuracy gains over the supervised learning baseline by 7.1\%, and over recent supervised state-of-the-art, CDS by 2.1\% on remote sensing Xview data.