DMRS: Long-tailed remote sensing recognition via semantic-aware mixing and diversity experts

Long-tailed class distributions pose a significant challenge in remote sensing scene recognition, where certain scene categories appear far less frequently than others. However, existing long-tailed learning approaches often overlook the unique spatial hierarchies and contextual semantic relationships inherent in remote sensing imagery, limiting their effectiveness in this domain. To address this, we propose Diversity-Mix Remote Sensing (DMRS), a foundation model-based framework designed for long-tailed remote sensing scene recognition. DMRS introduces two key innovations: (1) multi-low-rank adaptation diversity experts, which achieves balanced classification by specializing different experts for different regions of the class distribution, and (2) a semantic-aware mixing strategy, which incorporates textual semantic information typically unused in traditional classification to enhance perception across diverse remote sensing scenes. Extensive experiments on NWPU-RESISC45 and RSD46-WHU datasets demonstrate the effectiveness of DMRS, achieving 6.7% and 2.0% improvements in overall accuracy, respectively, while significantly enhancing the recognition of tail classes. These results highlight the potential of DMRS in tackling long-tail challenges in remote sensing scene classification. The data and codes used in the study are detailed in: https://github.com/wyfhbb/DMRS . • Proposed DMRS framework enhances long-tailed remote sensing recognition with semantic-aware mixing and diversity experts. • Novel Diversity LowRank Adaptation (D-LoRA) employs specialized experts to focus on different regions of the class distribution. • Mixing with Semantic Similarity Strategy (MixSSS) enriches feature space of tail classes by fully utilizing category name semantic information. • Maintained robust performance on tail classes while requiring significantly reduced training time compared to existing methods.