Multi-Proxy Quality-Aware Learning and Classifier Calibration for Few-Shot Incremental Fine-Grained Remote Sensing Classification

Few-shot class-incremental learning (FSCIL) enables models to continuously learn new classes with limited samples while retaining knowledge of previously seen classes. Despite significant progress for natural scenes, FSCIL for fine grained remote sensing (RS) image classification remains under explored. Existing FSCIL methods are difficult to adapt to RS characteristics: multiple subtle features, uneven image quality distributions, and background clutter. To bridge these gaps, we propose REC, a three-stage decoupled learning-based approach that concurrently refines REpresentation and calibrates biased Classifiers. At the Base training, a multi-proxy classification strategy allocates several learnable proxies as classifier heads to model each class. Optimal-transport matching aligns retained local features with their optimal proxies, thereby capturing subtle visual distinctions. A quality-aware re-weighting scheme leverages the Euclidean norm of backbone embeddings as the internal quality indicator, prioritizing high-quality samples and suppressing low-quality ones. During the Pseudo-incremental stage, pseudo sessions are generated by sampling support, query sets from base data, where the way of support set is progressively expanded to emulate the increasing proportion of novel classes. A light-weight classifier calibration module is trained to contextualize all existing classifiers and samples for classifier debiasing. During real incremental sessions, only the calibration module adds negligible overhead relative to the vanilla decoupled baseline. REC achieves SOTA on fine-grained RS datasets, Fair1m and ShipRsImageNet, with consistent cross-domain generalization to natural-scene benchmarks, MiniImageNet and Cub200.