CRGNet

CRGNet performs weakly supervised semantic segmentation of very high-resolution (VHR) remote sensing images by expanding sparse point-level annotations with a consistency-regularized region-growing approach.

Key Features:

• Region-Growing Mechanism: An iterative region-growing algorithm expands sparse point-level annotations by selecting unlabeled pixels with high confidence.
• Consistency Regularization Strategy: Consistency regularization minimizes discrepancies between a base classifier supervised by original sparse annotations and an expanded classifier trained on region-grown annotations to mitigate error propagation.
• Dual Classifier System: A dual-classifier architecture (base classifier and expanded classifier) jointly refines segmentation and controls the fidelity of annotation expansion.

Scientific Applications:

• Land Cover Classification: Enables land cover classification on VHR remote sensing imagery with reduced pixel-wise annotation requirements.
• Urban Planning: Produces detailed semantic segmentation of urban features from VHR imagery to support urban planning analyses.
• Environmental Monitoring: Facilitates environmental monitoring by segmenting high-resolution land cover and land-use features from VHR imagery.

Methodology:

CRGNet starts from sparse point-level annotations, iteratively expands annotations via a region-growing algorithm that selects high-confidence unlabeled pixels, and applies consistency regularization by minimizing discrepancies between a base classifier supervised by sparse annotations and an expanded classifier trained on the region-grown annotations.