Toward Open-World Remote Sensing Imagery Interpretation: Past, present, and future

Artificial intelligence solutions, especially those based on deep learning, have swept through the realm of remote sensing image understanding over the past decade. Despite remarkable research advancements, there remains a gap between state-of-the-art techniques and application requirements. Neural models trained under conventional paradigms typically demonstrate limited capabilities in rapid generalization to new tasks, accurate identification of unseen categories, effective engagement in continuous learning, and adequate adaptation to distributional differences. Therefore, developing generic and scalable imagery interpretation approaches tailored for open-world scenarios is crucial. This article provides the first systematic and comprehensive review of key visual tasks within the remote sensing domain, encompassing few/zero-shot learning, open-set recognition, incremental learning, and domain adaptation/generalization. The anticipated objectives, evaluation protocols, and representative works for each task are thoroughly discussed, particularly highlighting their unique contributions to open-world image understanding. The review spans nearly 300 publications, with a strong focus on those from the past three years (2022-2024). Furthermore, this article conducts an in-depth analysis of the challenges facing the field, both at the level of individual visual tasks and within the broader context of open-world topics. Potential coping strategies and promising research directions, such as foundation models, multimodality, and domain knowledge, are also outlined to shed new insights on future endeavors in the geoscience and remote sensing community.