Visible red-tide patches in nearshore waters are often small, fragmented, and difficult to resolve using satellite ocean-color imagery because of limited spatial resolution, cloud cover, mixed pixels, and adjacency effects. Unmanned aerial vehicles (UAVs) provide very-high-resolution observations, but consumer-grade RGB cameras lack the near-infrared and red-edge bands commonly used in bloom detection. To address this limitation, we developed HABD-Net, a lightweight encoder–decoder network for fine-scale delineation of visible red-tide patches from UAV RGB imagery. The model integrates a ConvNeXt-based Algal Bloom Enhancement Module, which combines multi-scale depthwise convolutions with channel and spatial recalibration, and an Index-Guided Self-Attention module that uses the red-difference index, I_R=2R-G-B, as a weak spectral–spatial prior.

The primary dataset consisted of UAV RGB images acquired during a Noctiluca scintillans bloom in Pingtan, Fujian, China. Data were split at the original-image level, and 21 full-resolution images were independently evaluated using overlapping sliding-window inference and probability-map reconstruction. HABD-Net was compared with six representative segmentation models and further tested through cross-scene transfer on a smaller Prorocentrum donghaiense dataset from Shacheng Harbor, Ningde.

HABD-Net achieved a precision of 91.25%, recall of 89.20%, F1-score of 89.85%, and MIoU of 89.42%, while requiring only 15.81 million parameters and 2.02 GMACs. Attention-guided index integration provided a better balance between target coverage and false-positive suppression than direct index concatenation. Transfer learning increased Ningde MIoU from 78.77% to 82.25%. These results demonstrate the potential of HABD-Net for accurate and efficient UAV-based mapping of visible red-tide patches in optically complex coastal waters.