Class-Incremental Learning for Remote Sensing Scene Classification via Stable Diffusion-Based Data Regeneration

Remote sensing scene classification aims to semantically identify land-cover scenes from satellite or aerial imagery, serving as one of the fundamental tasks in remote sensing interpretation. With the continuous growth in the scale and complexity of remote sensing data, data distributions evolve dynamically, e.g., new categories emerge over time. Thus, continual learning (CL) for remote sensing scene classification has become increasingly important. Among existing CL approaches, experience replay methods achieve the most competitive performance; however, directly applying them to high-resolution and wide-coverage remote sensing imagery incurs substantial storage overhead and raises concerns about data security and privacy. To address these issues, we propose a diffusion-based generative replay scheme that eliminates the need to store real samples while significantly reducing storage requirements. Specifically, we employ low-rank adaptation (LoRA) to fine-tune the stable diffusion (SD) model for each sequential task, enabling it to adapt to evolve data distributions and generate representative samples from prior tasks for efficient replay. Furthermore, to enhance the quality of replayed data, we introduce a data selection mechanism that leverages the previously trained classifier to filter generated samples. Experimental results on multiple remote sensing datasets demonstrate that our proposed method can achieve the superior performance.