Akademik Veri Yönetim Sistemi
Journal of Computational and Applied Mathematics, cilt.487, 2026 (SCI-Expanded, Scopus)
High-resolution remote sensing semantic segmentation is inherently affected by mixed pixels, inter-class spectral similarity, and ill-defined boundaries, which often lead to fragmented predictions and unreliable confidence in transition regions. To explicitly represent and regulate this ambiguity within an encoder-decoder framework, this study introduces T2F-UNet, an Interval Type-2 Fuzzy Uncertainty-Guided U-Net that embeds a learnable interval Type-2 fuzzy gate into U-Net skip connections. The proposed gate estimates a pixel-wise uncertainty width map and modulates encoder features before their fusion with the decoder, thereby strengthening boundary aware feature learning and improving robustness in transition areas. raining is guided by a width-map–based objective that integrates uncertainty-weighted cross-entropy, multiclass Dice loss, and an uncertainty regularization term. This formulation is designed to emphasize difficult and ambiguous pixels while preventing uncontrolled inflation of uncertainty across homogeneous regions. Experiments are conducted under a consistent training protocol using a ResNeXt-50 backbone and 256 × 256 patches on two complementary benchmarks: DeepGlobe land-cover segmentation and MBRSC Dubai semantic segmentation. Across strong baselines (Vanilla U-Net, DeepLabV3+, PSPNet, FPN) and a Type-1 fuzzy variant, T2F-UNet delivers the best overall performance, achieving 0.7211 mIoU, 0.8206 F1-score, and 0.8354 overall accuracy on the DeepGlobe dataset, while attaining 0.7940 mIoU, 0.8815 F1-score, and 0.8842 accuracy on the Dubai dataset. In addition to quantitative improvements, the resulting uncertainty estimates are primarily localized along semantic boundaries and structurally complex regions, yielding interpretable confidence cues that are consistent with segmentation errors and support more reliable downstream geospatial analysis.