Semi-Supervised Segmentation and Boundary Accuracy in Aerial Mapping Imagery: A Model Audit

Abstract

The rapid proliferation of high resolution aerial mapping imagery has necessitated the development of automated, scalable solutions for land cover classification, infrastructure monitoring, and environmental assessment. While deep learning has revolutionized image segmentation, the reliance on massive fully annotated datasets poses a severe bottleneck, driving the adoption of semi supervised learning techniques. This paper conducts a comprehensive model audit of semi supervised segmentation frameworks applied to aerial imagery, with a specialized focus on boundary accuracy. Unlike standard natural images, aerial maps feature complex, densely packed objects with intricate geometries where precise boundary delineation is critical for practical applications such as cadastral mapping and urban planning. Through a systematic evaluation methodology, this research assesses how state of the art semi supervised paradigms, including consistency regularization and pseudo labeling, perform in capturing sharp morphological edges compared to their fully supervised counterparts. The audit reveals that while semi supervised models achieve comparable regional overlap metrics, they frequently exhibit significant degradation in boundary localization, often suffering from over smoothing and structural hallucination. By analyzing the interplay between unlabeled data integration and spatial precision, this study provides actionable insights into the limitations of current architectures and proposes conceptual pathways for boundary aware semi supervised training protocols.