Each of the 13 final class names in the dataset is a diagnosed disease state (or healthy state), not merely observable symptoms. Examples include Healthy, Bacterial Spot, Shot Hole, Yellow Leaf, and Powdery Mildew. The “Yellow Leaf” category is based on nutrient deficiency diseases (e.g., iron or nitrogen deficiency) diagnosed by tissue symptom patterns and expert interpretation. The name “Powdery Mildew” is used to describe a fungus disease, confirmed by direct expert inspection. Visual cues (e.g., coloration, lesions) were used in early-stage classification, with final class determination relying on validation by farm experts.

Table 3 summarizes the key parameters recorded during the image acquisition process, including class, date, location, temperature, weather conditions, and the devices used. Images of the four selected medicinal plant species were captured between November 01, 2024, and January 05, 2025, primarily at the National Botanical Garden in Mirpur, Dhaka, Bangladesh. The temperature during collection varied between 22 °C to 33 °C, while weather conditions ranged from sunny to cloudy, ensuring a wide range of natural lighting scenarios. To introduce diversity in resolution and perspective, images were captured using four different mobile devices. These factors collectively contribute to the dataset’s robustness, making it suitable for training machine learning models under varied real-world conditions.

The experimental design for the development of the medicinal leaf dataset was carefully planned to ensure high-quality, diverse, and reliable data suitable for machine learning applications. Images of four medicinal plant leaf types were captured using four different mobile cameras, introducing variability in lighting conditions, viewing angles, and image quality. The image collection took place at the National Botanical Garden, Mirpur-2, Dhaka (Latitude: 23.8127° N, Longitude: 90.3476° E) between November 01, 2024, and January 05, 2025.

All collected images were manually sorted into five distinct disease classes: Healthy, Bacterial Spot, Shot Hole, Powdery Mildew, and Yellow Leaf. The classification was validated by agricultural experts to ensure labeling accuracy and dataset integrity. To further increase dataset volume and variability, six data augmentation techniques—including rotation, zooming, flipping and brightness adjustment—were applied. All images were then resized to a uniform resolution of 512×512 pixels, ensuring compatibility across a wide range of machine learning and deep learning models.

In the model generation phase, a deep learning model such as MobileNetV2 was trained on the dataset to classify the different leaf conditions effectively. The final stage involved model evaluation, where performance metrics such as accuracy, precision, recall and F1 score were calculated to assess classification efficiency. This comprehensive workflow demonstrates how an AI-driven system can be leveraged for the early detection and classification of leaf diseases, contributing to more sustainable agricultural practices and improved medicinal plant health monitoring. Fig. 2 illustrates sample image classifications produced by the trained model.

The leaf images were captured using four different mobile phone cameras: Apple iPhone 12 Pro Max, Samsung Galaxy A13, iQOO Neo 9 and Vivo V27. These devices were chosen for their high-resolution imaging capabilities, which introduced diversity in terms of camera specifications, such as megapixel resolution, aperture size, and image dimensions. This variability ensures the inclusion of real-world inconsistencies in image data, such as changes in lighting, clarity, and focus—factors critical for evaluating the robustness of image processing and machine learning algorithms. By incorporating images from different devices, the dataset better simulates real-life conditions, making it a valuable resource for research in image enhancement, noise resilience, and generalization in classification tasks. Fig. 3 presents the flowchart of the entire dataset preparation and classification process, while Table 4 outlines the detailed technical specifications of each camera used, including the camera resolution (MP), aperture, and original image dimensions.

To ensure the development of a reliable expert system using machine learning and deep learning models, a carefully structured preprocessing pipeline was established. This pipeline consists of several crucial stages designed to maintain data quality, increase variability, and support effective model training. The following steps outline the entire process from data collection to classification:

Original Data: Containing raw images captured in the botanical garden (dimension 512×512).

Augmented Data: Containing the transformed images created through augmentation to improve model performance.

The dataset is systematically organized into two primary directories: Original Images and Augmented Images. Each directory contains images from four distinct leaf species, further categorized into five disease classes—Healthy, Bacterial Spot, Shot Hole, Powdery Mildew, and Yellow Leaf. The Original Images directory comprises raw, unprocessed photographs that have been directly sorted according to their respective disease types. This serves as the baseline dataset. In contrast, the Augmented Images directory includes images that have undergone preprocessing operations such as resizing and a variety of augmentation techniques, including rotation, flipping, zooming and brightness adjustments to enhance model training robustness. A schematic overview of this hierarchical organization is presented in Fig. 4, which visually illustrates the structured arrangement of leaf images by species and disease category.

To ensure the reliability and scientific accuracy of the medicinal plant leaf dataset, the annotation and validation process was conducted under expert supervision. Professor Dr. M. A. Rahim, Head of the Department of Agricultural Science, and Dr. ATM Majharul Mannan, Assistant Professor of the same department at Daffodil International University (DIU), Dhaka, Bangladesh, led this process.

The process of annotation went as follows:

To validate statistically, we conducted an inter-annotator agreement test on a random subset of 1000 images. The process yielded a Cohen's Kappa of 0.91, a high level of agreement and reliability of annotators.

With their extensive experience in plant disease identification and classification, the dataset underwent a rigorous validation protocol.

The annotation process involved several key steps:

In addition, the annotation of every image in a disease category was not decided solely by symptom patterns. Rather, expert validation ensured visual features matched accepted diagnostic criteria. For diseases like Powdery Mildew, the annotation was based on fungal growth patterns verified using plant pathology resources. For Yellow Leaf, the annotation is consistent with a diagnosis of nutrient deficiency disorder rather than broad yellowing symptoms. Where necessary, our agronomists consulted laboratory or secondary tissue condition references to confirm labeling accuracy.

To support transparency, a certification of validation has been included as a supplemental file, highlighting the scientific diligence behind the classification process. This expert-driven data annotation approach has resulted in a high-quality, dependable dataset, ideal for training and evaluating machine learning models for early plant disease detection and classification.

The initial model testing involved training a MobileNetV2 model to categorize images into 13 distinct classes within the dataset. The model was evaluated using a confusion matrix, which resulted in a 13 × 13 matrix, as shown in Fig. 5. Additionally, key performance metrics, including accuracy, precision, recall, and F1 score, were computed. The detailed performance results for each class are presented in Table 5.

From Table 5, the MobileNetV2 model performed well across all 13 classes, with high accuracy, precision, recall, and F1 scores. The accuracy for most classes ranged between 91 % and 94 %, indicating strong classification performance. Precision and recall were consistently high, with values above 0.90 for most classes, suggesting that the model was good at both identifying correct classes and minimizing false positives/negatives. The F1 score, which balances precision and recall, was also strong, typically ranging from 0.93 to 0.96. Overall, the model demonstrated excellent performance in classifying images from the dataset.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.