Invariant-feature segmentation unlocks robust large-area mapping of mariculture

Accurate, large-area mapping of mariculture infrastructure is essential for science-based marine spatial planning, yet most existing approaches exhibit limited transferability beyond small, homogeneous sites because they rely on scene-specific spectral thresholds or deep-learning models that require extensive, high-quality labels. Here we present an object-based, two-level refined-segmentation approach that exploits four domain-invariant features including grey-value contrast, regularity, rectangular fit, aggregation (GRRA) to detect floating cages, long-line rafts, and related structures in medium-resolution satellite imagery. Level 1 applies adaptive local thresholding (Otsu) to generate initial candidate objects; Level 2 sharpens boundaries with edge-guided refinement (Canny) and an iterative edge-threshold scheme that self-tunes to local background conditions. Tested on Landsat-8 scenes spanning four major Chinese mariculture provinces (Liaoning, Shandong, Jiangsu, and Fujian), the approach achieved scene-level F-measures of 0.92-0.98 and outperformed a manually tuned OBIA baseline (0.63-0.95), without any scene-specific parameter retuning. Because the features encode intrinsic layout and morphological traits of human-designed farms rather than absolute water color, the algorithm demonstrated robust performance under heterogeneous conditions, including turbidity fronts, sea-ice interference, and sun-glint. The proposed approach is computationally efficient and scalable for updating mariculture inventories, and the GRRA framework has potential applicability to extracting other semi-regular coastal infrastructure (e.g., pond aquaculture, floating solar arrays) in dynamic marine environments.