Spatially-adaptive dynamic dictionary learning for fine-grained remote sensing image segmentation

Dynamic dictionary learning provides explicit class-level semantic prototypes for remote sensing image segmentation. However, most existing methods generate one global dictionary for the whole image through global feature aggregation, which overlooks the fact that the same class may have different local appearances under shadows, scale changes, seasonal variation, or background clutter. To address this limitation, we propose Spatially-Adaptive Dynamic Dictionary Learning (SA-DDL). Instead of following the conventional one-image-one-dictionary design, SA-DDL constructs a dense Dictionary Field in which each local region is assigned its own class embeddings. The framework contains a Local Context Modulator for topology-preserving dictionary generation and a Region-Aware Interaction Decoder for aligning pixel features with their corresponding local dictionary atoms. Experiments on LoveDA, UAVid, Potsdam, and Vaihingen show that SA-DDL consistently improves segmentation accuracy over strong recent baselines, while additional ablation, parameter sensitivity, and perturbation studies verify the effectiveness and robustness of the proposed spatial adaptation mechanism.