AgriVision-Apricot

Description

A multi class Image Dataset for Deep Learning-Based Apricot Pathology and Stress Classification

Steps to reproduce

======================================= AGRIVISION-APRICOT: STEPS TO REPRODUCE ======================================= 1. IMAGE ACQUISITION & DATA COLLECTION -------------------------------------- A. Location & Sampling: High-resolution digital imagery of apricot trees (Prunus armeniaca) was collected across various agricultural orchards and field environments under natural outdoor lighting conditions. B. Imaging Hardware: Images were captured using mobile camera sensors at variable distances (10 cm to 1 m) from target plant organs. C. Environmental Variability: Samples were gathered under diverse weather conditions (cloudy, sunny, direct sunlight, and partial shade) to capture variations in natural illumination and background complexity. 2. DIAGNOSIS & CLASS ASSIGNMENT (EXPERT ANNOTATION) --------------------------------------------------- A. Pathology Verification: Each image was visually inspected and categorized into one of nine target classes: 1. brown_rot 2. dieback 3. gumosis_and_galls 4. healthy 5. other_disorder_e.g_water_stress 6. powdery_mildew 7. shot_hole_fruit 8. shot_hole_leaves 9. wilting B. Symptom Identification: Classification was assigned based on characteristic visual markers for biotic diseases (e.g., fungal lesions, shot holes, powdery mildew coating) and abiotic stress symptoms (e.g., foliar drooping, water deficiency symptoms). C. Validation: Ambiguous images with overlapping symptoms or low visibility were removed or cross-verified with domain experts in plant pathology to ensure high label accuracy. 3. DATA CLEANING & CURATION --------------------------- A. Quality Control: Out-of-focus, heavily distorted, or duplicate frames were filtered out to preserve dataset quality. B. Class Organization: Images were sorted into single-label subdirectories corresponding to their primary pathology/condition for straightforward loading into deep learning image classification pipelines (e.g., PyTorch ImageFolder or TensorFlow image_dataset_from_directory). 4. REPRODUCTION & BENCHMARK PIPELINE ---------------- To reproduce baseline deep learning benchmarks using this dataset: 1. Load the dataset (base line or augmented as per need) directory into an image classification environment. 2. Apply a random stratified split (e.g., 70% Train, 15% Validation, 15% Test) across all 9 folders. 3. Preprocess images by resizing to model input dimensions (e.g., 224x224) and normalizing pixel values using standard ImageNet mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225]. 4. Train baseline convolutional or transformer architectures (e.g., ResNet-50, EfficientNet, ViT) using cross-entropy loss and standard data augmentations (flips, slight rotations, color jitter). 5. For effective generalization of the results, it is highly recommended to augment the dataset as per requirements. ======

Institutions

• National University of Computer and Emerging Sciences

  Islamabad, Islamabad

Categories

Licence