MangoFruitBD: A Bounding-Box Annotated Image Dataset for Detecting Healthy and Diseased Mango Fruits in Bangladeshi Orchards

MangoFruitBD is a bounding-box annotated RGB image dataset for detecting and localizing healthy and diseased mango fruits. It contains 1,310 JPEG images and 2,127 fruit-level annotations across four classes: Alternaria, Anthracnose, Healthy, and Scab. Of these, 1,286 images contain annotated fruit and 24 are fruit-free orchard scenes with empty YOLO label files as negative samples. The dataset also includes 128 multi-fruit images; each fruit is annotated independently, and “Mixed” is an image category rather than an object class. The annotation distribution is 1,073 Healthy, 435 Anthracnose, 317 Scab, and 302 Alternaria bounding boxes. Labels are stored in YOLO text format using class index and normalized bounding-box centre coordinates, width, and height. A total of 1,072 images were captured under orchard conditions in Natore and Rajshahi districts, Bangladesh, during 3–10 June 2025 and 1–5 June 2026. Images were acquired using Realme X2 and Samsung SM-E225F smartphones and retain natural variation in illumination, viewing angle, distance, object scale, background clutter, overlapping fruit, and partial occlusion. Images were preserved at original resolution without cropping, colour correction, or uniform resizing. The Alternaria category was collected differently because affected fruit could not be reliably photographed while attached to trees. Therefore, 155 retained original Alternaria images were captured against a controlled, non-orchard background. A further 83 images were generated from these source images using Python and OpenCV through rotation, horizontal flipping, brightness adjustment, and noise addition. The final Alternaria category contains 238 images, all associated with the controlled-background setting: 155 originals and 83 augmented derivatives. No augmentation was applied to other categories. The exact numerical augmentation parameters could not be verified from the retained records and are therefore not reported. Images with severe blur, poor focus, duplication, or inconclusive symptoms were removed. Approximately 20% of images per class were independently reviewed by Md. Rokonuzzaman, Sub-Assistant Agriculture Officer. Bounding boxes were manually created in CVAT and checked for placement, class assignment, missing files, and filename correspondence. The dataset was split at image level into training, validation, and test subsets using an 80:10:10 ratio with random seed 42, producing 1,048 training, 131 validation, and 131 test images. Images and labels are arranged in parallel YOLO-compatible directories, with a data.yaml file defining paths and class names. MangoFruitBD supports fruit detection, disease localization, model benchmarking, orchard monitoring, and postharvest inspection. Users should consider class imbalance, geographic concentration, resolution variation, and the controlled-background nature of the Alternaria category. The dataset is released under CC BY-NC 3.0.