Class Bias Correction Matters: A Class-Incremental Learning Framework for Remote Sensing Scene Classification

Abstract

Most existing deep learning models for remote sensing scene classification (RSSC) adopt offline learning paradigm, which are trained on closed datasets and fail to dynamically update with new class data. Currently, class-incremental learning (CIL) allows models to learn new classes while retaining discrimination of old ones. However, most CIL approaches aim to overcome catastrophic forgetting by employing techniques such as exemplarmemory and knowledge distillation, while ignoring the prediction bias caused by imbalanced datasets, where old classes retain fewer samples than new ones. Moreover, they do not adequately account for the multilevel semantic structure and multiscale feature information inherent in remote sensing images (RSIs). To address these issues, we propose an effective CIL framework for RSSC, named class bias correction network (CBCNet). Specifically, a cross-dimensional and interaction-aware attention mechanism (CIAM) is designed to incorporate channel, position, and direction-aware information in feature maps, enabling the model to highlight informative regions within RSIs. Next, a contextual information fusion module (CIFM) is proposed to explore the correlations among multilevel features and enhance representation quality through their fusion. In addition, the designed taskwise classifier head decoupling mechanism (TCDM) imposes a constraint to mitigate the prediction bias toward new classes, and enhances model’s discrimination among all seen classes. Finally, a multilevel integrated knowledge distillation module (MKDM) is developed to ensure comprehensive knowledge transfer, empowering the model to maintain critical representations in feature space and make well-informed decisions in output probability space. Experiments on five open datasets demonstrate the outperformance and robustness of our method.