AvocadoPest3
Description
The article presents an image dataset of avocado (Persea americana) leaves affected by three of the main pests impacting its cultivation: the persea mite (Oligonychus perseae), the red spider mite (Oligonychus punicae), and the avocado leafminer (Caloptilia perseae). The original dataset consists of 1,932 images captured in situ across various geographic regions of Mexico (Morelos and Guerrero), using multiple mobile devices to ensure model robustness to hardware variability. Following a curation and cropping process, 2,033 images of individual leaves were obtained and categorized into nine classes representing three severity levels (initial, intermediate, and advanced) for each pest, with validation by phytopathology experts. To mitigate class imbalance and improve algorithm generalization, data augmentation techniques were applied, resulting in a final collection of 4,871 images. The dataset is designed for training, validation, and testing of deep learning architectures to enable quantitative, automated estimation of foliar damage in avocado crops.
Files
Steps to reproduce
Data collection: 4,871 images of avocado leaves captured in situ across Morelos and Guerrero (Mexico) using four mobile devices (iPhone 14 Plus, OPPO Reno 7, Huawei P9 Lite, ZTE 9047) under natural light between 9:00 a.m. and 3:00 p.m., with multiple viewpoints (top-down, side, oblique). Individualization: Raw field captures containing multiple leaves were manually cropped to isolate single-leaf instances, yielding 2,033 individual leaf images across three pests (persea mite, red spider mite, leafminer). Labeling: Images were classified in Roboflow into nine classes representing three severity levels (Initial, Intermediate, Advanced) for each pest, with expert phytopathology validation (class counts pre-augmentation: CrystalMite 240/240/227; LeafMiner 239/223/238; RedMite 201/201/224). Preprocessing: Images were standardized to 224 × 224 pixels (resize, stretch), with auto-orientation and EXIF orientation stripped to ensure consistent input for model training. Augmentation: Field robustness was increased via random horizontal/vertical flips (p=0.5), 90° rotations (none/clockwise/counterclockwise), random crops (0–20% zoom-in), small-angle rotations (±15°), and photometric adjustments (brightness ±15%). This expanded the dataset by 139.6% to a final size of 4,871 images. Final dataset: 4,871 images at 224 × 224 pixels organized for image-level severity classification across three avocado pests (Persea mite, Red spider mite, Leafminer), with distribution by severity: Persea mite 1,697; Red spider mite 1,637; Leafminer 1,537.