A Dynamic Confidence-Aware Mask Fusion Framework for Aerial Image Segmentation Using U-Net and U-Net++

Abstract

Semantic segmentation of aerial and drone imagery remains a demanding task due to class imbalance, scale diversity, and the visual similarity between land-cover categories. Although encoder–decoder architectures such as U-Net and U-Net++ have demonstrated strong performance, reliance on a single model often limits generalization across heterogeneous datasets. This study introduces a Dynamic Confidence-Aware Mask Fusion (DCAMF) framework that reformulates model fusion as a pixel-wise, uncertainty-guided decision process. Instead of applying static averaging or conventional voting schemes, the proposed approach integrates heterogeneous U-Net and U-Net++ models equipped with multiple backbone encoders and assigns adaptive weights based on entropy-derived confidence measures. The framework was evaluated on two multi-class aerial datasets with differing levels of semantic complexity. Experimental findings indicate consistent improvements in mean Intersection over Union and overall pixel accuracy compared with individual architectures and traditional ensemble strategies. The gains were particularly evident in boundary regions and visually ambiguous areas, where model disagreement is common. By incorporating uncertainty modeling directly into mask-level aggregation, DCAMF enhances robustness without introducing architectural modifications or excessive computational overhead. The results demonstrate that confidence-aware decision fusion offers a practical and scalable pathway for improving segmentation reliability in real-world aerial imaging applications. Received: 15 November 2025 | Revised: 27 May 2026 | Accepted: 19 June 2026 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement Data sharing is not applicable to this article as no new data were created or analyzed in this study. Author Contribution Statement Ahmad Qasim AlHamad: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization. Ahmad Mousa Odat: Validation. Abdullah Ahmad Al Sokkar: Supervision. Mohammed Abdallah Otair: Conceptualization, Methodology, Resources, Writing – review & editing. Suhier Nadi Odah: Formal analysis. Omar Hussain Tarawneh: Project administration.