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

Area of Science:

• Remote Sensing
• Marine Ecology
• Geographic Information Systems (GIS)

Background:

• Accurate mapping of mariculture infrastructure is crucial for effective marine spatial planning.
• Existing methods often lack transferability due to reliance on scene-specific parameters or extensive labeled data.
• There is a need for robust, scalable techniques for large-area mariculture monitoring.

Purpose of the Study:

• To develop and validate an object-based, two-level refined-segmentation approach for mapping diverse mariculture structures.
• To exploit domain-invariant features for robust detection in medium-resolution satellite imagery.
• To assess the approach's performance across different environmental conditions and geographic locations.

Main Methods:

• An object-based, two-level refined-segmentation approach utilizing four domain-invariant features: grey-value contrast, regularity, rectangular fit, and aggregation (GRRA).
• Level 1 segmentation uses adaptive local thresholding (Otsu) for initial object detection.
• Level 2 refines boundaries using edge-guided refinement (Canny) and an iterative edge-threshold scheme.

Main Results:

• Achieved high scene-level F-measures (0.92-0.98) across four major Chinese mariculture provinces.
• Outperformed a manually tuned OBIA baseline (0.63-0.95) without scene-specific parameter retuning.
• Demonstrated robust performance under heterogeneous conditions like turbidity, sea-ice, and sun-glint.

Conclusions:

• The GRRA-based approach provides accurate and transferable mapping of mariculture infrastructure.
• The method is computationally efficient, scalable, and adaptable to various coastal environments.
• The GRRA framework shows potential for mapping other semi-regular coastal infrastructure.