SP-DDLNet: A Detail-Driven Dynamic Dictionary Learning Network for Green Tide Detection From Remote Sensing Imagery
Abstract
Accurate green tide segmentation requires simultaneously preserving the continuity of large-span ribbon-like structures and delineating fine-grained fragmented regions. Existing segmentation models may struggle to balance these two requirements because coarse semantic representations can weaken local structural details, while local features alone provide insufficient contextual information. To address this issue, we propose SP-DDLNet, a detail-driven dynamic dictionary learning framework. A ribbon-aware geometric aggregator (RAGA) captures axis-aligned long-range dependencies through strip pooling to enhance the spatial continuity of elongated green tide structures. A detail-aware dynamic modulator (DADM) then jointly integrates deep semantic features and shallow structural cues to generate scene-adaptive semantic tokens from a learnable dictionary. These tokens are further refined with spatial features through the bidirectional interaction in the detail-aware dynamic decoder (DADD), and a hypernetwork generates instance-adaptive reconstruction kernels for final prediction. Experiments on the moderate resolution imaging spectroradiometer (MODIS), (250 m) and Gaofen-1 (GF-1), (16 m) datasets achieve intersection over union (IoU)/ $F1$ scores of 0.7696/0.8714 and 0.8293/0.9067, respectively, demonstrating consistent improvements in segmentation accuracy across different spatial resolutions.