BIMIM: Band-Independent Masked Image Modeling With Transformer for Multispectral Satellite Imagery

Self-supervised learning (SSL) offers a promising solution to reduce reliance on labeled data. Among SSL approaches, Masked Image Modeling (MIM) has demonstrated significant potential in remote sensing applications such as scene classification and semantic segmentation, owing to its ability to capture pixel-level details. However, existing MIM frameworks, originally designed for natural images, struggle to adapt to the spectral-spatial characteristics of multi-spectral satellite imagery. While recent studies have introduced spectral-enhanced MIM SSL methods, most rely on band-group embedding, which imposes constraints on band utilization flexibility in downstream fine-tuning tasks and limits the granularity of spectral feature learning. To address these challenges, this study proposes Band-Independent Masked Image Modeling (BIMIM) with Transformer, a novel self-supervised learning framework specifically designed for multi-spectral satellite imagery. BIMIM not only enables finer band-specific spectral feature extraction, allowing for more effective capture of subtle spectral variations, but also introduces spatially random masking at the single-band level, facilitating more efficient inter-band feature learning. Extensive experiments on publicly available remote sensing datasets demonstrate that BIMIM achieves state-of-the-art performance in downstream tasks such as scene classification and semantic segmentation. This study provides a new perspective on SSL for multi-spectral remote sensing, paving the way for more effective spectral-spatial feature extraction and adaptation in self-supervised learning frameworks.