Performance Analysis of Various Deep Transfer Learning Models for Bacterial Blight Disease Detection and Classification in Pomegranate Fruits

Agriculture remains a vital pillar of the Indian economy, according to the FAO (Food and Agriculture Organization of the United Nations) in their 2020 report, around 40% of the global population relies on agriculture for their livelihoods. However, fruit diseases pose significant threats to food security (Rehman et al., 2021), with 40% of fruit loss attributed to pests and diseases and 25-30% lost during post-harvest stages (Technology Information, Forecasting and Assessment Council). The spread of diseases throughout farms can result in substantial losses in both quality and quantity of fruits. Tragically, crop failure contributes to 11.2% of all suicides in India (Hossain et al., 2020), emphasizing the urgent need for effective disease detection and prevention methods. Precise and prompt identification of fruit diseases is crucial for farmers, consumers and policymakers to develop strategies for trade-offs, pricing, storage and addressing food shortages.

Pomegranate fruit production is widespread globally, with several countries contributing significantly to its cultivation (Singh et al., 2023). Iran is the world’s leading producer of pomegranates. Pomegranates are deeply rooted in Persian culture and Iran’s climate is well-suited for their cultivation. India is another major producer of pomegranates, with states like Maharashtra andhra Pradesh and Karnataka leading the production (Naseer et al., 2024). Indian pomegranates are known for their sweet taste and vibrant color (Gawade et al., 2018). When pomegranate fruit is attacked by diseases, it can have various effects on the fruit itself, the plant and the overall yield of the orchard. Diseases can directly impact the yield of pomegranate orchards by causing premature fruit drop, quality decline, increased susceptibility to pests, spread to other trees, reducing fruit size, or even killing the fruit before it reaches maturity (Wakhare et al., 2023). There are several diseases, like Anthracnose (K Jayalakshmi et al., 2015), Alternaria (Aloi et al., 2021), Fruit Rot, Bacterial Blight (D.M. Sharath et al., 2019), Cercospora (Gk, Ravichandra et al., 2023), Fusarium wilt (AlZubi, A. A.et_al2024) etc., which can have various effects on their appearance, quality and marketability.

However, current manual detection methods are time-consuming and costly, requiring continuous expert monitoring. Laboratory testing, while effective, demands optimal conditions and specialized knowledge. Hence, prioritizing automatic early detection and prevention measures is crucial to mitigate disease spread, safeguard food security and bolster economic development. This approach not only saves resources but also protects food donors and enhances the nation’s prosperity. This research aims to evaluate different CNN-based architectures to improve the accuracy of pomegranate disease detection and reduce losses. The proposed system aids farmers in identifying diseases and suggests suitable treatments and preventive actions to mitigate losses.
 
Review of literature
 
Pal Arunangshu et al., (2023) presented AgriD et al. framework for plant disease detection and severity classification. AgriDet outperforms previous models in accuracy and is validated through statistical analysis for metrics including accuracy, specificity and sensitivity. Khatawkar et al., (2023), provide a system that utilizes a machine-learning approach to identify and categorize two major diseases affecting pomegranates, namely Fruit Rot and Scab, during their early stages. The dataset comprises 1000 images, including 250 each of Fruit rot and Scab diseases and 500 healthy fruit images. Classifier accuracy, evaluated via confusion matrix, reaches 83% for disease detection. Nirmal et al. (2022) suggested a framework utilizing a Support Vector Machine to classify healthy and diseased leaf images. The dataset collected from the Mendeley portal contains a total of 559 images, with obtained accuracy of 95.53%.

Nirgude et al. (2021) employed pre-trained models ResNet50, ResNet18 and InceptionV3 for detecting and classifying pomegranate diseases. The dataset created by the authors contains 1493 images. ResNet50 achieved 97.92% accuracy, ResNet18 achieved 87.5% accuracy and Inception-V3 achieved 78.75% accuracy. Javeriya Syeda et al.  (2021) utilized the Faster R-CNN deep learning model for detecting and classifying pomegranate fruit diseases. Bhange Manisha et al., (2015), introduced a web-based application facilitating farmers in identifying fruit diseases via image uploads. Specifically targeting pomegranate fruit, the system utilizes a pre-existing dataset for comparison. Authors continuously expand the dataset to enhance system accuracy, achieving an 82% detection accuracy for Bacterial Blight in pomegranates. Sachin B. Jadhav et al., (2021) employed pre-trained CNN models to identify plant diseases. Their dataset comprised 649 images of healthy and diseased plants, focusing on three soybean diseases: Frogeye leaf spot, Bacterial blight and Septoria brown spot. Achieving accuracies of 98.75% and 96.25% for the respective CNN models. Thorat, A.,et_al(2023) introduced a pomegranate fruit disease classification system employing a pre-trained VGG16 model in conjunction with a transfer learning approach.

In the research outlined in Yan et al. (2023), a deep learning model employing a transfer learning approach was introduced to detect Fusarium wilt disease in banana crop leaves. Their model achieved an accuracy of 98% and an F-1 score of 98% when evaluated on a dataset comprising 600 samples of both diseased and healthy banana leaf images. Geetharamani et al. (2019) suggested that transfer learning has been applied in a variety of fields, including software defect prediction, sentiment analysis and plant classification. In essence, transfer learning operates by incorporating the expertise of an earlier trained CNN model into a newly created CNN model intended for a particular task Shaha et al. (2018). In a transfer learning-based plant disease detection task Mukti et al. (2019) the authors assessed the four previously trained models. ResNet-50 was discovered to be the most accurate model, achieving an accuracy of 99.80%. Srivastava et al. (2023) present a deep learning framework for detecting and classifying plant leaf diseases, employing five DCNN models. MobileNetV2 achieves the highest accuracy of 98.9% for dataset1, while DenseNet121 achieves 99.9% accuracy for the Cherry Dataset. Pakruddin et al. (2024) concluded that various techniques for pomegranate fruit disease detection and classification using machine learning and deep learning have been effectively consolidated.

All the experiments were conducted at the research center named Rashtreeya Vidyalaya College of Engineering Bangalore in the academic year 2023-2024.
 
Dataset used
       
The proposed pomegranate fruit disease dataset contains 5099 images of five categories Alternaria, Anthracnose (Chavan and Dhutraj et al., 2017), Bacterial Blight (Sharath et al., 2019), Cercospora and Healthy images. In the current study, a total of 2400 images portraying both healthy and bacterial blight diseases were utilized. These images were partitioned into training and testing sets, with 80% allocated for training purposes and 20% for testing. The dataset consists of five folders named by related medical conditions. The original images were captured in JPG format, featuring a 1:1 aspect ratio and dimensions of 3120 ´3120 pixels. These pomegranate images were obtained within both sunny and cloudy environments on a farm (Pakruddin et al., 2024). Fig 1 showcases a selection of images from the dataset put forth in this study.


 
Proposed model architecture
 
The author created the proposed PomeNetV1 model, a lightweight, straightforward 12-layer CNN model, to categorize the health and diseases in the proposed dataset images. The suggested model does not make use of transfer learning, in contrast to other models that are put out in this study. Just 6 layers 2 fully-connected dense layers and 4 convolutional layers have trainable weights and biases. The proposed model was created specifically to battle overfitting by including a smaller number of layers, in each layer less number of filters. There are 3 dropout layers are also added for not recognizing insignificant patterns. Each convolutional layer uses a 3´3 kernel size, while each max-pooling layer employs a 2´2 pool size. Feature extraction is conducted from the initial 8 layers of the model, while the final 4 layers are utilized for generating classification results. The model is trained from scratch since no transfer learning is used.


The 1st layer receives the input images of size 220´ 220 and applies 64 3´3 filters to them. A ReLU activation function is then applied to this layer’s output. The output of the 1st layer is fed into the 2nd layer, which uses a ReLU activation function and 64 filters with a 3´3 size. The 2nd layer’s output is sent into a max-pooling layer with a 2´2 filter size to reduce the dimensions of the data, 108´108 image size and to extract important features from the input. For regularization, the output of the max-pooling layer is fed into the dropout layer with a probability of 0.2 per cent. The 3rd layer receives the input images of size 106´106 and applies 128 3´3 filters to them. A ReLU activation function is then applied to this layer’s output. The output of the 3rd layer is fed into the 4th layer, which uses a ReLU activation function and 128 filters with a 3´3 size. The 4th layer’s output is sent into a max-pooling layer with a 2´2 filter size to reduce the dimensions of the data, 52´52 image size and to extract important features from the input. For regularization, the output of the max-pooling layer is fed into the dropout layer with a probability of 0.2 percent. After being flattened into a vector, the dropout layer’s output is passed into two fully connected layers, each containing 512 and 2 neurons. A ReLU activation function and a dropout layer for regularization with a learning rate of 0.001 come after the first fully connected layer. A probability distribution across the 2 potential classes is produced by the softmax layer, which is the last layer. 
 
Transfer learning and fine tuning
 
The tensor flow open-source platform provides 38 pre-trained deep-learning models based on Keras. These models can be used for various applications, including transfer learning, optimization, forecasting, feature extraction, image recognition, natural language processing and handwriting recognition. This technique, widely used in deep learning, leverages pre-trained models’ weights (initially trained on ImageNet for image classification) as a starting point and fine-tunes them on new data. This approach greatly reduces the computational resources and time needed for training, particularly when labeled data is scarce. Fine-tuning involves unfreezing some or all layers of a pre-trained model and retraining it with new data (Yan  et al., 2023). In our study, we froze the top layers of all pre-trained models and applied transfer learning only to the fully connected layers. We then modified the architecture by replacing the 1000 neurons in the output layer with two neurons to suit our health and disease classification objective.

Pre- trained models
This work presents pre-trained deep-learning models for ImageNet based on Keras. Table 1 summarizes the essential characteristics of the pre-trained models.


 
Experimental Setup
 
For all tests, Tensor Flow version 2.14.0 and Keras version 2.14.0 are utilized. All of the code is written and executed using Google Colab. Table 2 provides particular hardware parameters for the experiment.



An essential part of training deep learning models is optimizing hyperparameters. Hyperparameters are predefined configuration settings crucial for training, not derived from the data itself. Optimizing these parameters can substantially influence the performance of our deep-learning model. Table 3. Summarizes some common hyperparameters used for pre-trained models and proposed models.



 In deep learning, data augmentation is frequently utilized as a strategy to improve the performance and generalization of models, especially when datasets are constrained. The primary objective of data augmentation is to apply several transformations to the existing data, artificially enlarging the size of the training dataset. Rotation, flipping, zooming, cropping and other basic data augmentation techniques are used in deep learning use cases. A few of the methods for pre-trained models and the proposed model are shown in Table 4.




 The network architecture utilized in all pre-trained models except the proposed PomeNetV1 model where transfer learning or fine-tuning is applied is shown in Fig 2 and Fig 3 describes the proposed PomeNetV1 model architecture.




The two primary components of the network architecture consist of the fully connected layers and the convolutional layers. Two Dense layers and a flattened layer make up the classifier. With 25 nodes, the first dense layer has “ReLU” as its activation function. With 2 nodes, the second dense layer has “softmax” as its activation function. However, section 4.1 provides the network architecture for the proposed model.

Due to the unavailability of a standard benchmark pomegranate fruit disease dataset, we have collected images in different places from real fields for dataset preparation, based on the symptoms of the pomegranate fruit it is categorized into five different diseases including healthy. A total of 5099 images were collected for experimentation. The data augmentation technique is used for training and testing deep learning models with different parameters. The training process involved the following steps.
1. Pre-trained models being loaded
2. Rebuilding the final three layers using the transfer learning technique to complete a new recognition task.
3. With the proposed dataset, train the model.
4. Verifying the outcome of the performance.
Fig 3 describe the proposed system architecture. Table 5 illustrates the comparative study of various models in pomegranate fruit disease classification.


 

The accuracy and loss given in Table 6, are the performance of the ten pre-trained models using the transfer learning approach and the proposed PomeNetV1 model without the transfer learning approach.



The Confusion Matrix and ROC curve for the Proposed PomeNetV1 Model are depicted in Fig 4 and 5, respectively.




From Fig 4, it’s evident that the proposed model correctly classified a total of 235 and 245 images, with only 5 images being misclassified. Fig 6, 7 and 8 depict training curves for the proposed PomeNetV1 model alongside two other models.





To provide a general understanding of the training process, these three models have been chosen because the proposed PomeNetV1 model is trained entirely from scratch, its initial training and validation losses started at relatively higher values, gradually decreased and eventually stabilized over the subsequent epochs. In contrast, because the VGG16 and ResNet50V2-based models already had a trained convolutional base, they both started with substantially lower loss values. The classification report for the proposed PomeNetV1 model with precision, recall, f1 score and accuracy is described in Table 7.


 

To classify pomegranate fruit diseases based on images, this research offered 10 different pre-trained deep learning-based models as well as a PomeNetV1 model. Except for the suggested model, which uses a pre-trained feature extractor for feature extraction and a modified classifier for final classification, Transfer learning or fine-tuning is beneficial for each of these models. In addition to these ten models, a straightforward sequential 12-layer PomeNetV1 model has been offered. The studies have been conducted using a publicly available proposed dataset of images related to pomegranate fruit disease. There are five classes in the dataset. Such a huge number of pomegranate fruit disease classifications was not the basis of any prior research. A novel assignment for identifying pomegranate fruit surface flaws is the classification of 2 types in this article. Prior studies have categorized them into more manageable categories, such as normal and abnormal but not with specific diseases. All pre-trained models produced the best classification accuracy of 99% for 50 epochs with transfer learning for pomegranate fruit health and disease image classification. Conversely, the proposed PomeNetV1 model exhibited outstanding performance with an accuracy of 99.80% over 50 epochs, despite its simple structure and compact size. This is mostly because of its feature extractor, which is made up of just four convolutional layers and is incredibly lightweight. Our future work is to evaluate a PomeNetV2 model on a proposed dataset for each disease detection and classification of the pomegranate fruits.
 

Public, commercial, or nonprofit funding agencies did not provide a particular grant for this study.

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.