Continual Barlow Twins: Continual Self-Supervised Learning for Remote Sensing Semantic Segmentation

In the field of earth observation (EO), continual learning (CL) algorithms have been proposed to deal with large datasets by decomposing them into several subsets and processing incrementally. The majority these assume that data are, first, coming from a single source, second, fully labeled. Real-world EO are instead characterized heterogeneity (e.g., aerial, satellite, or drone scenarios), for most part they unlabeled, meaning can be exploited only through emerging self-supervised (SSL) paradigm. For reasons, in this article, we present new algorithm merging SSL CL remote sensing applications call Barlow twins. It combines advantages one simplest self-supervision techniques, i.e., twins, elastic weight consolidation method avoid catastrophic forgetting. addition, evaluate approach on highly heterogeneous dataset, showing effectiveness strategy novel combination three almost non-overlapping domains (airborne Potsdam, satellite US3D, unmanned aerial vehicle semantic segmentation dataset), crucial downstream task EO, segmentation. Encouraging results show superiority setting, creating an incremental effective pretrained feature extractor, based ResNet50, without need relying complete availability all data, valuable saving time resources.