SIGNATURE WORK
CONFERENCE & EXHIBITION 2023

Remote sensing imagery provides a wealth of data that can be used by decision- and policy-makers to mitigate climate change consequences. Self-supervised learning offers the possibility of building meaningful representations from these images by learning to map the representations of similar images (artificially, geographically, temporally) closer to each other, while further from negative pairs. These representations can be used in turn in transfer learning to solve various downstream tasks with low training data and high accuracy which rivals and even surpasses that achieved by fully supervised methods. Given the increased ease of access to free, public remote sensing imagery, and the developments in self-supervised learning, this Signature Work Project explores the benefits of using self-supervision for limited labeled remote sensing data. We achieve this goal by creating a new remote sensing dataset and experimenting with pre-training various models with ImageNet, our model, or both. Our findings support previous research’s conclusions that self-supervision can effectively replace supervised techniques. We also find that SwAV pretrained on ImageNet only provides stellar performance on our three considered downstream tasks (solar PV detection, building segmentation, and crop field delineation) and that further pre-training on a remote sensing dataset improves performance only slightly, which does not warrant the additional effort in collecting the dataset and training on it too. These findings can inform practitioners of the best strategies to maximize the performance of self-supervised models for remote sensing data, which would allow for robust insights about the climate.