This article introduces Agri-Vision Bangladesh, a comprehensive, augmented image dataset designed to advance automated disease diagnosis in four economically vital agricultural crops: Bottle Gourd (Lagenaria siceraria), Zucchini (Cucurbita pepo), Papaya (Carica papaya), and Tomato (Solanum lycopersicum). Addressing the scarcity of region-specific agricultural data, a total of 5266 original images were acquired directly from diverse agricultural fields in Bangladesh using a SONY ALPHA 7 II full-frame camera under natural lighting conditions. The dataset encompasses 28 distinct classes, covering a wide spectrum of biotic stressors including viral (Mosaic Virus, Leaf Curl), fungal (Downy Mildew, Anthracnose, Alternaria Blight), bacterial (Bacterial Blight, Xanthomonas), and pest-induced damage (Insect Hole, White Spot), alongside Healthy samples. To ensure scientific reliability, each image underwent a rigorous two-stage validation process by senior agronomists. To tackle class imbalance and facilitate the training of data-intensive Deep Learning models, the dataset was expanded using a Python-based augmentation pipeline incorporating geometric transformations (rotation, flipping) and photometric adjustments (noise, brightness) resulting in a final repository of 28,000 images (5266 original and 22,734 augmented). All files are standardized to 512×512 pixels in JPG format. This expert-validated resource serves as a critical benchmark for developing robust computer vision algorithms (e.g., CNNs, Vision Transformers) for precision agriculture, enabling research into fine-grained classification, object detection, and cross-crop transfer learning in subtropical farming environments.
Billah et al. (Fri,) studied this question.