Intertidal Seagrass Mapping Using UAV Visible and Multispectral Imagery: A Comparative Semantic Segmentation Study with Explainability Analysis

Seagrass meadows are important blue carbon habitats, but their patchy distribution in intertidal zones makes accurate UAV mapping difficult under shallow water cover and complex sediment backgrounds. This study developed a fine-grained semantic segmentation framework with explainability analysis to improve intertidal seagrass extraction from high-resolution UAV visible and multispectral imagery. Exposed seagrass (ESG) and shallow-submerged seagrass (SSG) were mapped separately to represent two observable intertidal states. Visible bands, multispectral bands, and vegetation indices were used as model inputs. U-Net and DeepLabV3+ served as baseline models, while UPerNet-ConvNeXtV2-Tiny was tested under the same settings. Kernel SHAP and permutation importance were used to assess feature contributions. UPerNet-ConvNeXtV2-Tiny achieved the best performance, with an overall accuracy (ACC), mean Intersection over Union (mIoU), and F1 score of 97.45%, 94.63%, and 97.23%, respectively. It outperformed the baseline models in suppressing background interference, preserving patch morphology, and reducing omission errors in weak response and boundary areas, while demonstrating better cross-scenario applicability in independent test areas. Explainability analysis showed that model discrimination was mainly associated with red and green-related features, especially RGB-R, MS-R, MS-G, RGB-G, and NGRDI. ESG and SSG showed different feature dependence patterns, indicating that high-resolution UAV imagery can support accurate seagrass mapping and reveal spectral differences between intertidal seagrass states. These findings provide a practical framework for UAV-based intertidal seagrass mapping and monitoring and offer guidance for feature selection and model explainability analysis.