BD Guava Fruit and Leaf Disease Dataset

Bangladesh is a country located at the northern edge of the tropics. Geographically, it is bordered by the Himalayan mountain range to the north and the Bay of Bengal to the south, both of which strongly influence its climatic conditions. The country experiences a tropical monsoon climate, where the monsoon system is the dominant component of the annual weather pattern.
Bangladesh's tropical monsoon climate, characterized by high humidity, frequent rainfall, and persistent cloud cover, creates favorable conditions for the development and spread of guava diseases such as algal leaf spot, red rust, scab, and black mold. To support automated disease detection under realistic agricultural conditions, a guava disease dataset was collected during the monsoon season (June–September), with an average temperature of 27.71°C and relative humidity of 85.39%. Recent advances in deep learning, particularly YOLO-based architectures, have demonstrated strong performance in agricultural disease detection and motivate their application in this work.

In Version 4 of the dataset, label-level data and metadata information were further refined to improve dataset quality and usability. Each image is associated with structured metadata, including class labels, environmental attributes, and image-level details, ensuring better traceability and interpretability of the dataset. The dataset follows a standardized hierarchical organization in the format of train/validation/test splits, where each split contains class-wise directories for images and corresponding YOLO-format annotation files (.txt), such as train/class_name/images/class_name01.jpg and train/class_name/labels/class_name01.txt. Each annotation file stores bounding box information in YOLO format as class_id center_x center_y width height, for example 0 0.52 0.48 0.30 0.40, where 0 represents the class ID (e.g., algal leaf spot), 0.52 and 0.48 represent the normalized center coordinates of the bounding box, and 0.30 and 0.40 represent the normalized width and height respectively. This structured format ensures strict class-wise organization, maintains data consistency across all categories, and guarantees compatibility with deep learning frameworks by enabling efficient data loading, preprocessing, and training pipeline integration. The dataset is therefore organized in a well-defined format suitable for training, validation, and testing, ensuring reliable and reproducible experimental performance in guava disease detection tasks.
The Guava Disease Dataset is intended for a wide range of users, including agricultural researchers, machine learning researchers, data scientists, university students, agricultural extension officers, precision agriculture developers, software engineers, government agencies, AgriTech companies, and guava farmers.
GitHub Repository: https://github.com/shuvobasak4004/Guava-Multi-Model-Training-Code

孟加拉国作为番石榴生产大国，高度依赖季节性收获。然而，藻斑病、红锈病、疮痂病和黑霉病等病害严重影响番石榴产量与农户收益[4]。早期检测对于防止大规模感染、保障果品可持续生产至关重要。随着低成本算力与图像数据的可及性不断提升，深度学习为快速精准的病害识别提供了强有力的解决方案。

已有研究证明深度学习模型在同类农业领域具备极高应用潜力。例如，YOLOv8已成功应用于苹果园分割任务，在精度与推理速度上均优于Mask R-CNN，其精确率可达0.93、召回率达0.97，证实了其在实时农业应用中的鲁棒性[5]。类似地，在番茄分级任务中，YOLOv7在番茄品质检测与分类任务中实现了99.2%的准确率与99.4%的召回率，凸显了卷积神经网络（Convolutional Neural Network, CNN）在果品分析中的应用价值[6]。

受这些研究进展启发，本研究将YOLO与卷积神经网络模型应用于番石榴叶片与果实病害检测，以助力孟加拉国的精准农业发展。本研究将这些模型集成至Tkinter、Gradio等易用平台，旨在帮助农户与科研人员高效完成病害诊断，从而减少作物损失、提升果品质量。

版本2 红G
本数据集按类别划分文件夹，图像统一分辨率以保证一致性。每个类别包含充足样本以支撑模型的有效训练与评估。数据被划分为训练集、验证集与测试集三个子集，同时保留按类别组织的结构，以确保性能评估的平衡性与可靠性。

版本3 全类别集成
本数据集通过扩充样本多样性以提升模型泛化能力。通过对原始数据施加多种可控变换生成图像变体。最终数据集包含10个不同类别，且各类别样本量保持均衡。该数据以适配深度学习模型训练、验证与测试的结构化格式进行组织。

代码仓库：https://github.com/shuvobasak4004/Guava-Multi-Model-Training-Code