image-maks-spureous triplets

This dataset provides 31,441 image–mask–prediction triplets for binary semantic segmentation of paved urban roads from UAV orthophotography acquired over multiple municipalities in Colombia. Each triplet consists of three spatially aligned 256 × 256 pixel PNG files: a raw RGB UAV image tile, a ground-truth binary mask produced through manual digitization by two trained annotators in Global Mapper Pro, and a predicted mask generated by a U-Net convolutional neural network trained on the dataset. The dataset is partitioned into three standard splits: training (18,000 samples; 57.3%), validation (2,000 samples; 6.4%), and test (11,441 samples; 36.4%). All masks follow a binary encoding convention (road = 255, background = 0). The class distribution is near-balanced across all splits (mean foreground coverage ≈ 53.5%; FG/BG ratio ≈ 1.15), making the dataset directly compatible with standard binary cross-entropy training objectives without class reweighting. The U-Net baseline achieves an IoU of 0.8970 and a Dice Similarity Coefficient of 0.9447 on the test partition. The predicted masks are deliberately included as representative outputs exhibiting characteristic spurious geometric artifacts — irregular boundary noise and non-perpendicular edge deformations — that arise under standard training conditions in the absence of explicit geometric constraints. These predictions are intended as a structured foundation for research on geometry-aware post-processing and penalization strategies aimed at recovering topologically correct road vectors from deep learning outputs. The multi-municipality, variable-altitude acquisition strategy introduces natural scale variability across samples, supporting transfer learning and domain adaptation experiments.