DARC: Lightweight Density-Adaptive Label Relation Calibration for Multi-Label Remote Sensing Scene Classification

Multi-label remote sensing scene classification requires identifying multiple land-cover categories from a single high-resolution aerial image. Existing methods strengthen visual features or model label dependencies, yet they apply a fixed calibration strategy regardless of the underlying label-density regime, leading to over-prediction on dense scenes or under-correction on sparse scenes. We propose Density-Adaptive CDG Calibration (DARC), a lightweight framework that explicitly conditions calibration on dataset label density. DARC comprises three modules: (1) Label-density Driven Profile Selection (DDP) automatically routes the calibration path based on training-set density statistics; (2) Label-token Correlative-Discriminative Graph Mixing (CDM) injects both co-occurrence and exclusivity relations into label semantic tokens through positive and negative graph propagation; (3) Density-aware Gated Calibration (DCM) applies cardinality-controlled gating for dense labels and EMA-stabilized graph calibration for sparse labels. Experiments on AID-ML and UCM-ML demonstrate that DARC achieves 90.12% and 89.05% sample-F1, respectively, outperforming six competitive baselines including SFIN, ASL, C-Tran, ML-Decoder, SPIN, and Two-Way Loss by 1.65–3.87%, while introducing only 1.42% additional parameters. Cross-regime routing analysis confirms that no single fixed strategy matches DARC’s adaptive approach, and sensitivity analysis shows the routing is robust across a wide threshold range. Ablation studies validate the necessity of each component, and visualization analyses demonstrate that the learned label graphs capture interpretable semantic patterns.