Coarse mask-guided multitask learning for waterline extraction from optical coastal imagery

초록

This study presents a coarse mask-guided multitask learning framework for sand-water segmentation and waterline extraction from optical coastal imagery acquired across multiple platforms, including closed-circuit television systems, unmanned aerial vehicles (UAV), and satellite sensors. Instead of relying on laborintensive pixel-level annotations, the proposed approach employs a single coarse guiding mask and physically informed swash-zone water motion cues to jointly learn sand-water segmentation and waterline delineation in an annotation-efficient manner. A multitask learning formulation that combines segmentation and image reconstruction further enhances boundary accuracy under challenging coastal conditions. To robustly capture highly dynamic instantaneous waterlines affected by wave breaking and white foam, the framework exploits consecutive image frames rather than single-frame inputs. The proposed method was evaluated on optical coastal imagery from 45 viewpoints across 21 micro-or meso-tidal sandy beach sites along the east coast of Korea. Quantitative results demonstrate high segmentation performance, with average pixel accuracy of 99.64% on test datasets and 98.24% on blind test datasets. Geometric accuracy was further validated using distance-based evaluation against UAV real-time kinematic global navigation satellite system-derived ground-truth waterlines. On sandy beaches, the proposed method achieved mean perpendicular distance errors typically within 1-2 m, comparable to survey-grade measurements. While larger errors were observed in rocky and whitewash zones and for long-range oblique closed-circuit television (CCTV) views, the results highlight the complementary strengths of UAV and CCTV platforms for spatial accuracy and temporal continuity. Additional ablation and comparative experiments confirm the effectiveness of the coarse mask guidance, multitask learning strategy, and multi-frame inputs over conventional label-free approaches and state-of-the-art segmentation network and unsupervised segmentation approach as a baseline. Overall, the proposed coarse mask-guided multitask framework provides a practical, scalable, and platform-agnostic solution for automated instantaneous waterline extraction. By substantially reducing annotation requirements while maintaining high accuracy and robustness, this approach offers a promising tool for long-term coastal erosion monitoring, shoreline change analysis, and operational coastal management.

키워드