Dataset of Harumanis mango images for multi-modal maturity classification using ventral texture and size features.

This dataset contains 272 original Harumanis mango (Mangifera indica L.) images collected across three harvesting seasons (April to June 2023, 2024, 2025) from five orchards in Perlis, Malaysia. Images focus on the ventral surface region of the fruit, captured at 1443 × 2560 px against a black background using a Nikon D5600 DSLR. Labels were assigned to four maturity stages (Unripe, Semi-Ripe, Ripe, Rotten) by expert fruit growers following FAMA weekly counting standard. IMPORTANT - split and augmentation order: The 272 originals were split per class into training (97), validation (88), and test (87) subsets BEFORE any augmentation. Augmentation (rotation, flipping, brightness) was applied exclusively to the 97 training images, expanding them to 700. Validation and test subsets contain only original, non-augmented images, eliminating any risk of data leakage. Final dataset: 875 images total (700 train / 88 val / 87 test). Files: harumanis_phases_V2.zip - 272 original images (4 class folders) harumanis_phases_V2 augmented.zip - 875 images (train/val/test split folders) labels original 272 images.csv - class labels for all originals README.txt - full methodology details