Application of Image Processing and Transfer Learning for the Detection and Classification of Blemishes and Canker Diseases of Citrus

Citrus is a tropical and subtropical fruit crop, cultivated in more than 140 countries with 158.49 million metric tons annual output (FAOSTAT, 2024). Citrus hold an important role in agriculture and are generally consumed by everyone (Zhang et al., 2021). In Punjab (Pakistan) total fruit production is 30 %. Lemon, sweet orange, grapefruit and mandarin face various diseases like black spot and scab. These diseases not only lower fruit quality but also its yield. Traditional methods for diagnosis of disease is slow, less accurate and expensive but also depends upon expert intervention (Butt et al., 2025). The automated systems provide cost effective, consistent and efficient methods to mitigate the challenges of manual diagnosis. The suggested framework is compared to modern approaches, certifying its performance (Butt et al., 2024). The industry of agriculture has designate many deep learning (DP) models to tackle different challenges like fruit detection, plant leaf categorization, leaf and fruit disease detection and insect identification (Mamat et al., 2022). The approaches associated with DP, computer vision, and machine learning (ML) provide efficient solutions for classification of diseases and their detection. These digital technologies are integrated with precision agriculture for better management and diagnosis of disease through IoT based devising, drones and remote sensing which collect huge amount of information about plant health (Ayaz et al., 2019).

The application of DP, ML and different techniques of image processing has become critical for automated detection and classification of plant disease to minimize the need of manual inspection. These systems require minimum physical work, time and labor (Das et al., 2024). Supervised ML create models that categorize new data found on datasets. DL enables convolutional neural networks and hierarchical feature extraction highly effective for image analysis (Anwar et al., 2022). The computer techniques have made it uncomplicated to detect deviations in crops in real-time (Wang et al., 2017). These methods had significant success in recognizing and diagnosing diseases of plants. However, they have limitation in different task of image processing and image segmentation (Iqbal et al., 2018 and Singh et al., 2019), pattern recognition (Asghar et al., 2019), K-nearest neighbor method (Wang et al., 2020), feature extraction (Luaibi et al., 2021) and artificial neural network (Ali et al., 2021). Transfer learning (TL) is an effective approach for disease classification, uilizing pre-trained models such as ResNet and VGG16, which were developed to recognize large datasets images like ImageNet (Zhu et al., 2024). TL allows the application of previously trained models, including ResNet, GTM and VGG16 to categorize and identify citrus diseases as precisely as possible. It decreases the demands of huge data since the trained models already have a massive collection of attributes that are trained. The models do not require a fine tuning, making the training process faster (Hossen et al., 2025).

Another important use of such technology is the invention of mobile based technologies to detect and diagnose in real time. Farmers can receive real-time diagnosis by taking images of citrus plants with smartphones by integrating IP and TL algorithms into a mobile application (Nayak et al., 2023). A mobile application system has proposed for detection of disease in tomato. In this study scientist proposed a TL approach to fine tune convolutional neural network architectures like ResNet50, CGG19, SequeezeNet1.1 and DenseNet121. In this study highest accuracy achieved in Densenet121 model (99.85%) and this model was designed as an application in mobile phone for detection of disease in tomato. This program requires a high speed internet connection to perform diagnosis in 10 s (Nag et al., 2023). Another web based study was performed to diagnose disease with a validity of 93.91% (Asani et al., 2023). However, TL and IP for detection of disease in citrus have promising results, but also have several challenges. A great issue is the variability of disease symptoms and environmental factors (Qadri et al., 2025). By the creation of AI trained models leads to an increase in the farmers trust who gain a decision making process. The use of mobile based and drone based imaging will allow to track diseases in real time world.

A dataset comprising healthy and infected images of citrus fruits (canker and blemishes) was compiled from reliable agricultural databases, peer-reviewed publications, and smartphone photographs under natural lighting conditions. This study ensured a dataset with variations in lighting, angle, infection severity and fruit size that support better generalization of model.

The collected images were manually labelled into two categories: citrus blemishes and citrus canker (Figure 1). Accurate labelling was necessary for reducing label noise and ensuring the model could learn visual differences between disease types.

The preprocessing steps to enhance model learning and standardize input were:

Background reduction: The background elements were removed to focus on the region.

Resizing: The images were resized to meet the input requirements

Brightness normalization: The lighting variabilities were minimized to emphasized disease specific features.

The augmentation techniques (Figure 2) were used to enlarge the dataset size improve robustness were as following:

Rotation

Flipping

Zooming

Colour and contrast adjustments

A web based tool developed by Google (Saini et al., 2024) was used for model development and training (Figure 3).

Images were uploaded via local disk or webcam that ensured class completeness and balance.

The model was trained in the browser and images were labeled (Figure 4).

The model was assessed using a separate test dataset. The key metrics were:

Accuracy (correct predictions overall)

Precision (correct optimistic predictions out of total predicted positives)

Recall (correct optimistic predictions out of actual positives) (Malahina et al., 2024)

The trained model was utilized as a web-based link. The users can classify images of disease images by uploading a file or using a mobile camera. This makes the tool easily accessible to field officers, researchers and farmers, without software installation or technical expertise. It demonstrates a real-world application of AI in agriculture.

After training the model successfully, a confusion matrix (Figure 6) was generated to evaluate its performance in classifying two categories: citrus blemishes and citrus canker. Tit is used tool in classification tasks, detailing the differences between predicted and actual labels. It displays the number of instances in each category, helped to assess model misclassification and accuracy.

Table 1: Values of the confusion matrix

TP (True positive): The model correctly predicted citrus canker when the sample belonged to the citrus canker class.

TN (True Negative): The model correctly predicted citrus blemishes when the sample belonged to the citrus blemishes class.

FP (False Positive): The model incorrectly predicted citrus canker when the actual class was Citrus blemishes.

FN (False Negative): The model incorrectly predicted citrus blemishes when the actual class was citrus canker.

The confusion metrics revealed zero misclassification for FN and FP values. The model achieved perfect accuracy in the test dataset.

The performance metrics calculated were:

Accuracy

Definition: Total correct predictions proportion.

Formula:

Where

TP=225

TN=225

FP=0

FN=0 (Table 1)

Calculations:

Definition: How many were predicted as Citrus Canker?

Formula

Calculation

Definition: How many were correctly identified? Formula:

Calculation

Definition: Harmonic mean of recall and prcision.

Formula:

Calculation:

Figure 7 demonstrated the accuracy trend over 50 training epochs, representing both training accuracy (acc) and testing accuracy (test acc). The model quickly achieved a high accuracy of nearly 1.0 (100%), in the early stage and maintained its performance throughout. The minimum gap between training and test accuracy reflects better generalization on the test data. The consistent closeness of the curves suggests a stable learning process, free from underfitting or overfitting. It indicates that the model effectively learned to classify images and sustained its performance on unseen data.

Figure 8 demonstrated the loss values over 50 epochs, plotting both training and testing loss. Both losses were high and drop sharply within few epochs. Then they converged near zero (around 0.001) and remained stable throughout the training process. This decline indicated effective minimization of the error function. The minimum gap between the training and test loss curves suggests the model did not over fit. This convergence confirmed that the model was efficiently learned relevant patterns.

Our model has achieved a 100% recall, accuracy and F1-score in classifying citrus canker, with no FN or FP and TN and TP were 225. Similar results were observed by Odabas et al. (2024) thataccuracies ranging from 95% to 98% for Aphis spp. and Venturia inaequalis with minor misclassification in Monilinia laxa and Venturia inaequalis. Our model achieved perfect results and avoid overfitting. Malahina et al. (2024) observed trained class accuracies between 98 and their performance dropped to 35-96% on external dataset and highlighting generalization challenges. Ancheta et al. (2023) gained 98.17% accuracy for banana leaf diseases but they observed misclassifications between visually similar diseases, such as the black and yellow Sigatoka. Parate et al., (2023) observed 100% accuracy for leaves of guava, orange, and mango; however, strict metrics such as recall and precision were not reported, as shown in Figure 9. The evaluation was performed on a relatively small dataset of 300 images. In contrast of these studies, our observation not only achieved perfect accuracy but also integrate robust validation metrics. It included the F1 score, ensuring that the model avoids over fit and showed reliable generalization within tested dataset.

Previous studies showed the importance of accurate disease detetion methods. Kala et al. (2023) highlighted the crucial role of advanced technology in decreasing the effects of plant diseases on yields. Similarly, another study (Storey et al., 2022) highlight the significance of ML techniques for precise diagnosis of disease. These findings correlate with the outcomes of the present study, further validating the efficacy of utilizing a teachable machine for detection of disease.

• GTM performs optimally with small datasets (a few hundred images per class), whereas plant disease detection ideally requires thousands of high-quality images.
• With limited data, there is a risks of overfitting, where the model memorizes images rather than learning generalizable features, potentially reducing real-world accuracy.

• Complex backgrounds (e.g., soil, shadows) may mislead the model if not properly managed during training.
• Model performance can decline when training and test images differ in lighting, angles, or resolution.

The study reflects the efficacy of GTM for the detection and classification of citrus diseases, specifically blemishes and canker. Techniques such as image processing, data augmentation, and TL, the model achieved perfect performance metrics (100%). This indicates that the model has strong capability to distinguish between the two diseases with zero misclassifications and showed its potential as reliable alternatives to visual inspection methods. The model performance was affected by real-world variability in the training data. The dependency on GTM  imposes constraints on size of dataset and model complexity, which could impact scalability for broader applications in agriculture.

The authors would like to thank the Department of Plant Pathology, College of Agriculture, University of Sargodha, for providing facilities and support necessary to conduct this research.

The present research introduced the utilization of GTM with IP and TL for the classification and detection of citrus blemishes and canker. The model scored 100% accuracy, precision, recall and F1-score without over fitting. These findings indicate that AI-based tools using smartphone can provide a low cost and straightforward solution for citrus disease detection and precision farming.

Aqsa Batool: Trained the model, analyzed results and wrote first draft.

Hina Safdar: Supervised and guided the methodology and reviewed the manuscript.

Furqan-ur-Rehman: Provided technical support in data handling, validation and editing.

All authors read and approved the final version.

The authors declare that no generative AI or AI-assisted technologies were used in the writing, analysis, or preparation of this manuscript.

The authors have no conflict of interest.