A data-driven approach to turmeric disease detection: Dataset for plant condition classification

Abstract

Turmeric, Curcuma longa, is an economically and medicinally important crop. However, the crop has often suffered from diseases such as rhizome disease roots, leaf blotch, and dry conditions of leaves. The control of these diseases essentially requires early and accurate diagnosis to reduce losses and help farmers adopt sustainable farming methods. The conventional methods of diagnosis involve a visual examination of symptoms, which is laborious, subjective, and rather impossible in large areas. This paper proposes a new dataset consisting of 1037 originals and 4628 augmented images of turmeric plants representing five classes: healthy leaf, dry leaf, leaf blotch, rhizome disease roots, and rhizome healthy roots. The dataset was pre-processed to enhance its applicability to deep learning applications by resizing, cleaning, and augmenting the data through flipping, rotation, and brightness adjustment. The turmeric plant disease classification was conducted using the Inception-v3 model, attaining an accuracy of 97.36% with data augmentation, compared to 95.71% without augmentation. Some of the major key performance metrics are precision, recall, and F1-score, which establish the efficacy and robustness of the model. This work attempts to show the potential of AI-aided solutions towards precision farming and sustainable crop production in developing agriculture disease management. The publicly available dataset and the results obtained are expected to attract more research interest for innovations in AI-driven agriculture.

Specifications Table

Subject	Computer Science.
Specific subject area	Computer Vision, Agricultural Informatics, Plant Pathology, Artificial Intelligence, Plant Disease Detection, Crop Health Monitoring, Smart Farming
Type of data	Image.
Data collection	Manual collection of data for the research study has been done from different fields of Charpolisha in Jamalpur, from August 2024 to Janurary 2025. These collections were done under the observation of an agricultural expert. A total amount of 1073 raw images were collected from turmeric plantations showcasing a variety of symptoms, such as rhizome disease roots, rhizome healthy roots, dry leaves, healthy leaves, and leaf blotch. This robust dataset would be very important in the training of deep learning models in the early detection and effective management of diseases affecting turmeric plants to support sustainable agriculture.
Data source location	Town/City/Region: Charpolisha, Jamalpur Country: Bangladesh.
Data accessibility	Repository name: Mendeley Data. Data identification number: 10.17632/g46dvrcvwn.2 Direct URL to data: https://data.mendeley.com/datasets/g46dvrcvwn/2
Related research article	None.

1. Value of the Data

• •This dataset consists of several images regarding turmeric plant diseases, starting from the most prevalent to the rarest. Hence, it is quite valuable in terms of scientific research and agriculture. This will act as a stepping stone to further improve the plant pathology studies and enable early identification and classification of turmeric plant diseases. It is made openly available to maximize its impact, hence allowing collaborations amongst researchers and developers in similar areas such as image processing, plant disease classification, or related domains. It has been encouraging cross-discipline innovation that fosters several applications in agricultural management systems, disease prediction tools, and so on, helping both academic and industrial advances.
• •This dataset supports the training of CNNs and models based on Transformers for advanced deep learning architecture. Being a very rich labeled data set of turmeric plants, this dataset offers very robust models for the classification of diseases at very low computation expenses. Further, transfer learning steps involving a base model like VGG19, ResNet, or EfficientNet offer considerable performance enhancement and substantial shortening of training periods when using knowledge about features obtained by training other plant disease image databases.
• •This dataset offers a fantastic chance to assess AI models used in precision farming. Developers may test disease detection systems that are connected to mobile applications or Internet of Things sensors, allowing for automatic intervention and real-time monitoring in agricultural areas. The information can also enable prediction models that use machine learning and time-series analysis to anticipate disease outbreaks. Through focused interventions, these AI-powered solutions will assist farmers in reducing crop losses, enabling early diagnosis prior to the beginning of illness, implementing preventative measures, and increasing yields with fewer chemical treatments.
• •This dataset provides a platform for testing and refining a range of image processing methods, such as object recognition, edge detection, and picture segmentation, among others, to develop precise and computationally effective illness diagnosis algorithms—even under difficult circumstances. Practically speaking, these AI-powered technologies will assist farmers in more accurately identifying illnesses, lowering crop losses and raising yields. Additionally, by using focused interventions, these instruments can reduce the need for chemical treatments and guarantee sustainable agricultural methods. Thus, the creation of intelligent systems and innovation in the agricultural industry are stimulated by this dataset.

2. Background

Turmeric (Curcuma longa) is a crop of economic and medicinal importance. Like other crops, it is susceptible to a variety of diseases that seriously affect its productivity and quality. For the effective management of these diseases, detection, and identification should be done as early as possible to minimize the loss in agriculture and to preserve farming on a sustainable basis. Various studies have pointed out the catastrophic impact of plant diseases on crop yields, with estimates showing that global crop losses resulting from pests and pathogens lie between 20–40% annually [1].

Traditional plant disease diagnosis is based on expert visual inspections, which are labor-intensive, time-consuming, and full of subjective errors. This further worsens in large-scale agricultural fields. Advanced techniques, especially those with AI and deep learning, have shown huge potential to handle these challenges. AI-based solutions can provide fast, inexpensive, and accurate diagnosis of diseases, reducing the dependency on manual methods [2].

AI applications in agriculture are still facing critical challenges due to a lack of rich and diversified datasets for representation at different levels and manifestations of the disease in crops. Concerning previous work, Chenchupalli Chathurya et al. [3] concentrated solely on detecting whether a turmeric leaf is healthy or not, overlooking the critical need to identify root diseases, which were absent from their dataset. In contrast, R. Selvaraj et al. [4] employed a custom dataset featuring only four distinct classes of leaf diseases, enhancing the scope of their model. Meanwhile, Devisurya V. et al. [5] expanded their dataset to include images of turmeric plants afflicted by major diseases (three classes), such as leaf spot, leaf blotch, and rhizome rot, offering a more comprehensive view of the plant's health. By contrast, our dataset includes 1,073 images spread across five distinct classes: Leaf Blotch, Rhizome Disease Roots, Healthy Leaf, Dry Leaf, and Rhizome Healthy Roots. This expanded classification not only covers a broader spectrum of turmeric plant health issues, including both leaf and rhizome diseases, but also introduces categories like Dry Leaf and Rhizome Healthy Roots, which are crucial for distinguishing between healthy and affected plant parts. This comprehensive dataset, with its richer variety of plant conditions, enables more accurate, reliable, and versatile disease detection models, setting it apart from the narrower datasets used in previous studies. It also encourages transparency, reproducibility, and collaboration in research that catalyzes AI-powered agriculture in support of the very essence of sustainable farming.

3. Data Description

The dataset comprises 1073 high-resolution images of both healthy and diseased turmeric plants, categorized into five specific classes: Leaf Blotch Disease, Dry Leaf, Healthy Leaf, Rhizome Disease Roots and Rhizome Healthy Roots Disease. These images were collected manually during the late afternoon (3 to 5 p.m.) from turmeric fields in Charpolisha, Jamalpur, under the supervision of an agricultural expert. This time was specifically chosen to take advantage of optimal sunlight, and in favorable conditions, it also aligned with the golden hour in photography, providing soft and warm lighting that enhanced image quality. The collection period spanned from August 2024 to January 2025.All images were captured using the camera of an iPhone 14 Plus at a resolution of 800 × 800 pixels in JPG format. Each image is meticulously labeled with its respective disease class, making this dataset an invaluable resource for machine learning and image processing tasks related to the detection and classification of turmeric plant diseases. This comprehensive dataset is designed to aid researchers and practitioners in advancing agricultural disease management and developing robust diagnostic models for turmeric crop protection.

During the data collection process, several challenges were encountered, such as:

• 1.Capturing high-quality images of turmeric plants in specific areas posed several challenges. The variations in natural lighting, unpredictable weather, and shadows significantly impacted the clarity and consistency of the images. Additionally, working in a rural or remote location introduced logistical difficulties, such as transporting equipment and navigating poor road conditions, especially during unfavorable weather. These factors made the data collection process particularly demanding in this region.
• 2.The collaboration with agricultural experts was required for the proper labeling of the images, which was a very time-consuming and laborious task. The camera of the iPhone 14 Plus clicked high-resolution images but capturing manually, maintaining focus consistency, angles, and framing was time-consuming and prone to human error.

To address human annotation errors, several data cleaning steps were implemented. First, the labels for each image were reviewed multiple times by the agricultural expert to ensure consistency and accuracy, with any inconsistencies corrected through discussions and further validation. Additionally, a subset of images was cross-validated by the agricultural expert to minimize individual biases and mistakes. Lastly, a quality control process was introduced, where images flagged as potentially mislabeled were re-examined and re-annotated if necessary. These measures effectively reduced the impact of annotation errors and ensured the dataset's reliability for machine learning tasks.

These challenges, however, provide valuable insights and a solid foundation for the advancement of disease detection algorithms. Variations in image-taking conditions, such as fluctuating lighting and weather, present a real-world scenario for developing more robust and adaptable models that can perform well under varying environmental conditions. Additionally, the involvement of agricultural experts in annotating the images, despite being labor-intensive, ensures high-quality, expert-driven labels, which can be leveraged to train machine learning algorithms that can handle errors and ambiguities in data annotation. Addressing these inherent difficulties during data collection and annotation will ultimately improve the precision of automated systems for disease detection, advancing the field of agricultural diagnostics.

One field photo that is typical of the data gathering location is shown in Fig. 1.

Fig. 1.

The real turmeric field where images in the dataset were collected.

This dataset represents the state of real-world agricultural conditions, capturing the complexities and variability present in turmeric fields. By representing diverse environmental factors, it provides a robust framework for developing state-of-the-art disease detection algorithms. Such developments are important in building reliable automated diagnostic systems that can operate effectively across varying agricultural landscapes (Table 1).

Table 1.

Overview of diseases by turmeric plant class.

Name of Class	Description	Visualization
Leaf Blotch Disease	The disease symptoms manifest as small, oval, rectangular, or irregular brown spots on either side of the leaves, which gradually turn yellow or dark brown. The affected leaves also begin to yellow. In severe cases, the leaves may become scorched, reduction in rhizome yield [6]. The lesions vary in size, and some show advanced browning at the edges, indicating the progression of the disease. The yellowing around the spots further highlights the severity, with affected areas growing larger as the infection progresses
Dry Leaf	Includes plants with dried leaves or stems, usually due to environmental stress rather than specific pathogens [10]. The leaf appears completely dried out, with a curled and brittle texture. The color has turned brown, showing clear signs of dehydration typically caused by environmental conditions rather than any disease
Healthy Leaf	Represents turmeric plants free from any visible signs of disease [9]. This turmeric leaf is vibrant green, smooth, and without any discoloration, lesions, or irregularities. It appears strong and healthy, showcasing the plant's vitality and proper growth.
Rhizome Disease Roots	A fungal disease-causing decay in the roots, leading to stunted growth or potential plant death [7]. The roots appear discoloured, with a darker, unhealthy colour and a dry, shrivelled texture. There are visible areas of decay and fungal growth, with some parts of the roots appearing soft and deteriorated. The structure is compromised, and the roots show signs of rot, contributing to the decline of the plant's health
Rhizome Healthy Roots	Rhizome Healthy Roots refers to the portion of the turmeric plant's root system that is free from disease or damage, serving as a baseline for distinguishing healthy roots from those affected by conditions like rhizome rot [8]. These roots are firm, intact, and exhibit a pale yellowing colour with a smooth texture. There are no signs of decay or fungal growth, and the roots maintain their structural integrity and vibrant growth.

Agricultural research is advancing significantly with the integration of automation and machine learning, enhancing disease identification and control. The Turmeric Plant Disease Dataset plays a crucial role in this progress by providing high-quality images of various plant conditions, including healthy, dry, and diseased states such as leaf blotch and root rot. Expanding the dataset has notably improved model performance by increasing diversity, reducing overfitting, and enhancing generalization in real-world applications. To achieve this, we expanded our dataset to 1,073 original photographs and 4,628 augmented images, ensuring a comprehensive representation of each plant condition. The Statistics of the Turmeric Plant Dataset, represented in Table 2, presents the distribution of original and augmented images across different classes, explicitly detailing the original data count for each class and demonstrating how augmentation has strengthened model training.

Table 2.

Statistics of turmeric plant dataset.

Class Name	Number of Original Images	Number of Augmented Images
Dry Leaf	203	812
Healthy Leaf	197	985
Leaf Blotch	199	995
Rhizome Disease Roots	192	990
Rhizome Healthy Roots	282	846

This dataset will enable farm efficiency by facilitating automated disease detection and management. While conventional methods are time-consuming and error-prone, a machine learning-based approach can go through high-resolution images for early visual indications of disease expressions. This would speed up the process of not only identifying the diseases but also ensuring consistent results with accuracy, meeting local and global agricultural standards.

This dataset would definitely help develop various automatic grading systems that are quick in grading the quality of turmeric with a lot of efficiency. That allows large-scale and real-time monitoring of the diseases in the field of turmeric and manages those resources by including remote-sensing technologies using unmanned aerial vehicles called UAVs or satellite monitoring.

It fills an existing gap in real-world field-condition images which is essential in developing robust machine learning models that can perform well under natural environmental factors. This is also a contribution to the representation of turmeric plant diseases and supports the advancement of precision agriculture and sustainable farming practices.

4. Experimental Design, Materials and Methods

4.1. Experimental design

The profitability and production capacity of farmers is eventually impacted by the substantial influence that plant diseases have on agricultural yields. Computer-aided detection techniques are essential for accurately identifying plant diseases [11]. To assist agricultural specialists, this study focuses on diagnosing illnesses of Turmeric Plant Disease. To compile this dataset, we painstakingly gathered photos from Turmeric plantations, recording a range of Turmeric Plant Disease signs in diverse climates. Following data collection, we organized and cleaned the photos to improve their quality for analysis as part of the data preparation process. We used thorough preprocessing methods to increase the dataset's training appropriateness. Data augmentation techniques like flipping, rotation, and brightness modifications were applied to add more unpredictability to the dataset. After that, the dataset was split into two subsets: 20% for testing and 80% for training. The training subset was used to train deep learning models, and their performance was assessed using the test data. This procedure made sure that there were no noise or irregularities in the dataset. To establish a well-structured basis for building and evaluating models for precise Turmeric Plant Disease diagnosis, we lastly divided the dataset into training and testing sets. Additionally, the dataset includes classifications such as Leaf Blotch Disease, Root Disease, Healthy Roots, Dry Leaf, and Healthy Leaf, as shown in Fig. 2. By outlining the systematic process for identifying turmeric plant diseases, this dataset serves as a visual guide, fostering advancements in agricultural disease management and supporting crop health monitoring initiatives.

Fig. 2.

Workflow of the experimental design for the detection and classification of turmeric plant diseases.

4.2. Materials (camera specification)

We can vouch for the exceptional camera capabilities of the iPhone 14 Plus, which I used to take the photos. With a broad 120-degree field of view, the ultra-wide lens is perfect for photographing huge landscapes. Even in low light, crisp, clear images are guaranteed thanks to the 12 MP primary sensor and f/1.5 aperture. Deep Fusion and Night mode improve every picture in any kind of lighting. High contrast and brilliant colors are produced by the Photonic Engine. Every picture is a true depiction of quality and originality thanks to the front-facing 12 MP camera with Night mode, which also makes beautiful selfies.

4.3. Method

Following a few crucial procedures shown in Fig. 3, the raw photos gathered from the turmeric fields were treated during the pre-processing phase to get them ready for deep learning model training [12]. Before training, only high-quality samples were left after data cleaning was completed to eliminate noisy, unnecessary, or duplicate images. The photos were then downsized to standardize all of them to 800 × 800pixel resolution to preserve uniformity throughout the collection. Following data labeling, the photos were divided into five primary categories: dry leaf, root disease, healthy roots leaf blotch disease, and healthy leaf. Several data augmentations [13] methods were used to improve the dataset's resilience, including image shifting, flipping, rotation, zooming, adding noise, and altering brightness. By creating new training samples, these augmentations broadened the dataset's variety and decreased the possibility of overfitting when training the model. Above all, the augmentation strategies were essential for enhancing the deep learning model's generalization and resilience, guaranteeing its performance in real-world scenarios where image differences may arise. All things considered, the dataset was completely optimized by this pre-processing pipeline for further model training and assessment.

Fig. 3.

Pre-processing stages of the proposed deep learning model in the detection of turmeric plant diseases

4.4. Data augmentation

By generating new data samples from preexisting ones, data augmentation significantly improves model optimization and generalizability. When working with tiny datasets, this method efficiently expands the amount of training data that is accessible without requiring the gathering of extra data. All things considered, data augmentation is essential to raising the effectiveness and dependability of machine learning models in a variety of fields [14]. Data augmentation greatly improves deep learning models, especially those that are employed for visual object recognition tasks. Adding a variety of picture modifications to the training dataset enhances the model's generalization abilities and helps avoid overfitting. A variety of augmentation techniques were used in this study, such as flipping, rotating, zooming, shifting, adding noise, sharing, and brightness adjustment. By introducing diversity into the dataset, these techniques improve the model's capacity to classify turmeric plant illnesses in various circumstances, including changes in leaf orientation, lighting, and angles, and allow it to generalize effectively to unknown samples.

Table 3 provides a summary of the particular augmentation methods used, together with information on their ranges and characteristics. By applying these augmentations to the training dataset, the variety of the samples was greatly increased, which improved model performance.

Table 3.

Data augmentation techniques and parameters.

Techniques	Parameters	Ranges
Flipping	Direction	Horizontal
Rotation	Angle	${90}^0$ clockwise
Zoom	Zoom Range	0.8 to 1.2
Noise	Mean, Standard Deviation	0,25
Height Shifting	Shift amount	10% of image height
Brightness	Brightness Factor	0.5 to 1.5
Shearing	Shear Factor	0.1 to 0.5

Augmentation plays a critical role in improving the robustness and generalizability of a model. To achieve this, we applied several techniques that helped the model better adapt to various visual conditions encountered in real-world situations. Fig. 4 demonstrates the results of applying these techniques to images from different disease categories

Fig. 4.

Augmented images from the turmeric plant disease dataset.

Flipping, rotating, zooming, shifting, noise addition, shearing, and brightening are all applied to the original photos by the augmentation pipeline for various turmeric plant situations, including leaf blotch, dry leaves, rhizome roots, healthy roots, and healthy leaves, as seen in Fig. 4. These enhanced photos give a better understanding of the diversity that was included to make sure the model is capable of identifying and categorizing plant diseases in a variety of difficult situations. This augmentation technique is essential since it improves the model's accuracy and dependability in practical applications.

4.5. Model visualization

The Inception-v3 model, a variant of the Inception network, is well known for its effectiveness and precision in picture classification tasks [15]. To categorize turmeric plant illnesses, we carefully assessed the Inception-v3 architecture in this study. Images from four different classes, each of which represents a certain state of the turmeric plant, make up the dataset. We set aside 80% of the data for training and 20% for validation. We used data augmentation approaches to improve the training process, which broadened the training data's variety and enhanced the model's ability to generalize to new situations. A batch size of 32 was used to train the Inception-v3 model across 20 epochs. The learning rate was dynamically modified using a learning rate scheduler during training. The models' effectiveness was evaluated using a range of metrics, such as accuracy, precision, recall, and F1 score (Eqs. (1) to (4)) [16]. These metrics offer a comprehensive assessment of the models' performance and shed information on their multifaceted classification abilities. We also monitored the training process using training and validation datasets. The training metrics were visualized and analyzed using TensorBoard.

$$\text{Accuracy}=\frac{TruePositive+TrueNegative}{TruePositive+TrueNegative+FalsePositive+FalseNegative}$$ (1)

$$\text{Precision}=\frac{TruePositive}{TruePositive+FalsePositive}$$ (2)

$$\text{Recall}=\frac{TruePositive}{TruePositive+FalseNegative}$$ (3)

$$\mathrm{F}1\text{Score}=2\times \frac{Recall\times Precision}{Recall\pm Precision}$$ (4)

A confusion matrix is a matrix that shows the performance of a machine learning model on a collection of test data [17]. In our investigation, we used the confusion matrix to assess how well our model classified lemon leaf disease. The ground truth labels for each sickness category are shown graphically in the matrix along with the model's predictions.

The confusion matrix (Fig. 5) shows the model's high accuracy of 97.36% with augmentation, correctly classifying most classes like Dry Leaf and Healthy Leaf, with minor misclassifications between Rhizome Disease Roots and Rhizome Healthy Roots. In contrast, the confusion matrix (Fig. 6) shows the model's performance without augmentation, achieving 95.71% accuracy, with similar misclassifications but still strong overall performance.

Fig. 5.

Confusion matrix for Inception-v3 data with augmentation.

Fig. 6.

Confusion matrix for Inception-v3 data without augmentation.

Using the Inception-v3 model with augmentation, the classification achieved an accuracy of 97.36%, highlighting the model's strong performance and its potential for accurately diagnosing turmeric plant diseases across various agricultural conditions. Without augmentation, the model achieved an accuracy of 95.71%, still demonstrating solid reliability and effectiveness in disease classification. The classification reports with and without augmentation are provided in Table 4, Table 5, respectively, offering detailed metrics for each scenario (Fig. 7, Fig. 8).

Table 4.

)Model evaluation metrics classification report and accuracy for Inception-v3 (with augmentation).

Model	Class Name	Precision	Recall	F1-score	Accuracy
Inception-v3	Dry Leaf	0.99	1.00	1.00	97.36%
	Healthy Leaf	1.00	0.99	1.00
	Leaf Blotch	1.00	0.98	0.99
	Rhizome Disease Root	1.00	0.89	0.94
	Rhizome Healthy Root	0.88	1.00	0.94

Table 5.

Model evaluation metrics classification report and accuracy for Inception-v3 (without augmentation).

Model	Class Name	Precision	Recall	F1-score	Accuracy
Inception-v3	Dry Leaf	0.98	1.00	0.99	95.71%
	Healthy Leaf	1.00	1.00	1.00
	Leaf Blotch	1.00	0.97	0.99
	Rhizome Disease Root	1.00	0.78	0.88
	Rhizome Healthy Root	0.88	1.00	0.93

Fig. 7.

Accuracy & loss graph for Inception-v3 (with augmentation).

Fig. 8.

Accuracy & loss graph for Inception-v3 (without augmentation).

This study provides a comprehensive analysis of turmeric plant disease classification using advanced deep learning techniques. The Inception-v3 model was employed to identify various conditions affecting turmeric plants, utilizing a well-curated dataset. The results showcase the model's impressive performance, achieving an accuracy of 97.36% with augmentation. Its ability to accurately distinguish between different turmeric plant diseases demonstrates strong potential for real-world agricultural applications, facilitating early detection and effective prevention. Furthermore, the study's evaluation of performance metrics and the use of data augmentation techniques offer valuable insights for enhancing the model's effectiveness and interpretability.

Limitations

The Turmeric plant dataset is limited by its geographic focus, potentially reducing its generalizability to other regions with different environmental conditions. It may also suffer from class imbalances, which could affect model performance on underrepresented diseases. Additionally, the use of a single image capture device further limits diversity, and the absence of data on non-visible symptoms restricts its comprehensiveness for broader disease detection tasks.

Ethics Statement

None of the authors of this article have conducted research involving human or animal subjects. The datasets referenced in this article are publicly accessible, but adhering to proper citation guidelines is crucial.

CRediT Author Statement

A K M Fazlul Kobir Siam: Conceptualization, Methodology, Data curation, Writing. Md. Asraful Sharker Nirob:Conceptualization, Methodology, Visualization, Data curation, Writing. Prayma Bishshash: Conceptualization, Methodology, Writing. Md Assaduzzaman: Supervision, Data curation, Writing. Apurba Ghosh: Supervision, Writing – review & editing. Sheak Rashed Haider Noori: Writing – review.

Data Availability

• Mendeley DataTurmeric Plant Disease Dataset: Advancing AI for Agricultural Sustainability (Original data).