CherryLeaf-KG: Four-Class Cherry Leaf Disease Image Dataset with Uzbek Labels

Description

This dataset contains 400 cherry tree leaf images collected from active farms in Uch-Korgon village (Kadamzhay District, Batken Region, Kyrgyzstan) using an iPhone 14. Images were captured across three separate farms in July, shortly after the June cherry harvest, from mature trees of at least ten years old. The dataset is organized into four classes, named in Uzbek as used by local growers: Sarik (leaf yellowing / chlorosis, 100 images), Teshik (shot-hole disease with perforations, 100 images), Chirish (brown rot with decaying tissue, 100 images), and Soglom (healthy leaf, 100 images). All images were captured under real outdoor farm conditions, naturally including variations in illumination, leaf orientation, shadows, occlusions, and background clutter. Each image was manually annotated using makesense.ai, with a bounding box drawn around the central leaf. Images were then cropped and standardized into square format at 224x224 pixels. This dataset is designed as a benchmark for evaluating few-shot and zero-shot learning methods in real-world, resource-constrained agricultural settings, and serves as a cross-lingual test case due to its Uzbek-language class labels.

Steps to reproduce

1. Data Collection Images were collected on a single day in July, shortly after the June cherry harvest, across three farms in the village of Uch-Korgon (Kadamzhay District, Batken Region, Kyrgyzstan). All sampled trees were mature (at least 10 years old). Photos were taken outdoors under natural farm conditions using an iPhone 14 at a resolution of 960×1280 pixels. 2. Class Definition Four classes were defined using the terminology of local cherry growers: Sarik (yellowing), Teshik (shot-hole disease), Chirish (brown rot), and Soglom (healthy). Each class contains exactly 100 images. 3. Annotation Each image was manually inspected and the central leaf was localized using makesense.ai. A bounding box was drawn around the main leaf and exported as a CSV file. 4. Cropping and Preprocessing Using a Python script, each image was cropped according to its bounding box coordinates and standardized into a square image focused on a single leaf. All crops were then resized to 224×224 pixels for model input, using ImageNet mean and standard deviation normalization (mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225]). 5. Dataset Structure The final dataset is organized into four folders, one per class, each containing 100 preprocessed leaf images.

Institutions

• Eötvös Loránd University

  Budapest, Budapest
• Mawlana Bhashani Science and Technology University

  Dhaka Division, Dhaka

Categories

Licence