1. Introduction

Convolutional neural networks have been shown to be an outstanding tool for diagnosing and locating problems in stone fruits as well as in other crops by utilizing their leaves. Plant diseases affect the leaves, fruits, stems, roots, and overall quality and yield of the crop. As a consequence, there is a global deficit in fruit and vegetable consumption. Each year, crop diseases result in a 16% reduction in agricultural yield. The use of CNN designs for detecting plant diseases, including those affecting stone fruits, has been highlighted in various articles. The most modern methods for crop identification, feature fusion and extraction, training, data augmentation, image segmentation of various diseases of plants were inspected in these articles. The constraint of smaller datasets is mentioned as a major issue in over 80% of research publications ^{[1, 2]}. Rather than focusing on a small number of important classes, there should be a greater variety of stone fruit diseases ^{[3- 5]}. Therefore, the goal of this research is to look at how well CNN architectures like ResNet50 ^[6], Inceptionv3 ^[7], MobileNetV2 ^[8], and DenseNet ^[9] classify stone fruit leaf illnesses. Finding the model with the best accuracy rate for classifying stone fruit leaf diseases identification is the final aim of this research. The manual method of detecting plant diseases takes longer, is limited to certain areas, and is more prone to error by humans. As a result, the need for automatic illness detection methods has grown. This paper’s primary contributions are the tests conducted on collected data utilising four CNN architectures–MobileNetV2, DenseNet201, Inceptionv3, and ResNet50–that employ transfer learning. After that, a performance comparison is done to show which CNN model performs the best.

2. Literature Review

The relevant research work is reviewed in this part, the sorts of common diseases that affect stone fruits are included first, followed by different architectural configurations of CNN models, and lastly, research done by scientists to identify plant diseases using CNN models.

3. Methodology

3.2. Procedure

Firstly, images of plant leaves are taken. Colour, shape, and pattern of the leaf are extracted automatically in deep learning. On the basis of these attributes, deep learning algorithms are subsequently used for categorization. The entire process is mentioned below.

3.2.1 Raw data

RGB images are selected as an input. Images that did not clearly depict illness symptoms were excluded from the collection. Images of stone fruit leaves with a variety of bacterial, fungal, and viral infections are displayed in Fig. 1.

Fig. 1. Sample images of stone fruits leaves.

3.2.2 Image augmentation

The objective of augmentation is to raise the dataset’s variance while making sure that newly added data have significant additions ^[24]. The Keras DL framework is used for image augmentation. The following six types of augmentation options are employed in training: Rotation which arbitrarily rotate a picture from different perspectives. In brightness the model is fed images with varying brightnesses during training, which aids in its ability to adapt to changes in lighting. Shear turn the shearing angle of the image clockwise or anticlockwise. In zoom the provided image has multiple scales. In horizontal and vertical flips, the image’s axes are freely reversed in the horizontal and vertical flip modes.

3.2.3 Model training and classification

Based on the provided dataset, four pre-trained models-MobileNetV2, DenseNet201, InceptionV3, and ResNet50-were trained. Their classification accuracy was then evaluated. Fig. 2 provides a description of the experimentation process. All of the models underwent 40 epochs of training because, at that point, there would be no discernible improvement in validation loss. For DenseNet201, the fine-tuning time is −38 s (s)/epoch (Iterations), while for ResNet50, Inceptionv3, and MobileNetV2, it is 17 s/epoch and 14 s/epoch, respectively. Table 1 lists the hardware and software requirements for training CNN models, and also displays the CNN models’ experimental parameter values used in this analysis.

Fig. 2. Procedure of experiments.

Table 1. Hardware/Software requirements and Parameters used in study.

Software & Hardware Requirements
Configuration Item	Value
CPU	AMD Ryzen 5 5625U with Radeon Graphics 2.30 GHz
GPU	T4
Operating System	Windows 11 (x64-based processor)
RAM	12.67 GB
Disk	78.19 GB
Development Environment	TensorFlow with Keras (On Google Colab)
Programming Language	Python 3
Setup used in Experiment
Parameters	Value
Learning Rate	.0001
Drop out	0.3
Optimizer	Adam
Batch Size	32
Momentum	0.9
Epochs	40
Activation function	ReLu
Loss function	Categorical cross-entropy

4. Experimental Setup

This section presents the experiments findings based on the transfer learning of the pre-trained individual networks. The following experimental questions are anticipated to be addressed by the obtained results.

To answer the above-mentioned questions the subsequent performance metrics are taken into account:

5. Results

5.1. Answer to First Experimental Question “Which CNN Model Provides Better Accuracy in Detecting Stone Fruits Leaf Diseases?”

The performance of the four CNN models that use transfer learning is shown in this section. Table 2 displays the models’ performance in class-wise categorization. The results in Table 2 shows that ResNet50 consistently gets the highest F1-score, accuracy, precision, recall, and accuracy across most classes, especially for diseases of mangos and peaches. For example, ResNet50 outperformed the other models with a noteworthy accuracy of 96% for classifying healthy mangos and 96.2% for bacterial spot identification in peaches. ResNet50’s superior metrics across a variety of classes demonstrate its efficacy in precisely diagnosing crop diseases in stone fruits, even though all models performed well. With the y-axis representing accuracy and loss percentages and the x-axis representing the number of epochs, Fig. 3 displays the training & validation accuracy, training & validation loss versus epochs of the ResNet50 model and also shows the recognition accuracy, precision, recall and F1-score of individual model. Validation loss did not decrease from 0.27053 at the 40th epoch. Training accuracy was 96.4% and validation accuracy was 93.11% with this loss amount. The Resnet50 and Inceptionv3 models perform the best when taking into account each model’s precision values on the test dataset.

Table 2. Class-wise performance of pre-trained CNN models.

Class wise performance of CNN models
Model	Class	Accuracy	Precision	Recall	F1-Score
MobileNetV2	Mango_powdery_mildew	94.6%	94.0%	92.2%	93.2%
	Mango_gall_midge	91.6%	93.0%	95.2%	94.3%
	Mango_healthy	94%	95.3%	92.2%	92%
	Olive_aculus	93.8%	91%	92.2%	92.2%
	Olive_peacock	92%	92.8%	92.7%	91.7%
	Olive_healthy	93%	94.2%	91.4%	92.2%
	Peach_bacterial_spot	96.9%	96%	96.1%	96.2%
	Peach_rust	90.3%	87.2%	89.3%	89.2%
	Peach_healthy	92.3%	92.3%	92.4%	92.4%
DenseNet201	Mango_powdery_mildew	91.6%	92.6%	90.2%	89.4%
	Mango_gall_midge	91.8%	92.0%	93.7%	90.2%
	Mango_healthy	89%	88.2%	87.6%	89.7%
	Olive_aculus	92.4%	87.1%	87.4%	88.0%
	Olive_peacock	90.3%	90.8%	90%	90.8%
	Olive_healthy	92%	91.6%	89.3%	88.3%
	Peach_bacterial_spot	93%	94.6%	94.8%	93.9%
	Peach_rust	86.3%	87.5%	88.4%	89.0%
	Peach_healthy	91.2%	90.8%	91.9%	90.7%
InceptionV3	Mango_powdery_mildew	93.5%	94.0%	91.3%	90.2%
	Mango_gall_midge	93.6%	93.0%	94.2%	92.2%
	Mango_healthy	91%	89%	88.2%	89.2%
	Olive_aculus	90.3%	92.3%	91.2%	94.0%
	Olive_peacock	90.2%	92.8%	91%	90.3%
	Olive_healthy	93%	90.6%	90.4%	89.2%
	Peach_bacterial_spot	94%	95.5%	94.4%	95.3%
	Peach_rust	83.2%	86.5%	89.3%	89%
	Peach_healthy	90.2%	93.8%	92.9%	93.7%
ResNet50	Mango_powdery_mildew	95.6%	95.0%	94.2%	94.2%
	Mango_gall_midge	93.6%	93.0%	94.2%	93.2%
	Mango_healthy	96%	95%	95%	95%
	Olive_aculus	94.3%	95%	94.6%	94.0%
	Olive_peacock	93%	93.8%	92%	92.1%
	Olive_healthy	92%	93%	92.3%	92.1%
	Peach_bacterial_spot	96.2%	96.6%	96%	96%
	Peach_rust	89.2%	88%	87.2%	87%
	Peach_healthy	94.2%	93%	94.2%	94%

Fig. 3. (a) Accuracy vs epochs. (b) Loss vs epochs. (c) Performance metrics graphs of CNN models.

6. Discussion

In this study, we used transfer learning to conduct a thorough analysis of the outputs from four CNN models. We applied these models viz. MobileNetV2, DenseNet201, Inceptionv3, and ResNet50–to the nine classes of stone fruits and compared the outcomes as shown in Table 2. 36600 leaf photos of three stone fruits–mango, olive, and peach–are included in the dataset that was employed. Following the 70:30 split, we had 10980 images for testing and 25620 images for training. ResNet50, one of the CNN models, has the best classification outcomes for determining the illnesses of stone fruits’ leaves. In addition, Resnet50 is supported by ^[25] using either SVM or MobileNet. This is due to the fact that the fundamental concept of ResNet50 is based on the utilisation of shortcut connections. By avoiding intermediary layers, information can move straight from one layer to another. ResNet addresses the vanishing gradient issue that frequently befalls deep neural networks by adding residual blocks. This allows very deep networks to be trained without compromising on performance. Our research indicates that for smaller datasets, transfer learning of deep learning models offers good accuracy. Using the pre-trained model for training and feature extraction served as the foundation for the transfer learning techniques used in this study. Remarkably, as shown by analysis, the ResNet50 model outperformed previous state-of-the-art efforts. Table 3 compares the performance of various deep learning models on crop disease detection across different datasets and fruit types. Authors in ^{[1, 5]} focused on mango disease detection with smaller datasets, achieving accuracies of 89% and 88.46%, respectively, using AlexNet and a CNN-based model. Experiments in ^[21] utilized a larger UAV-acquired (Unmanned ariel vehicle) dataset for olive detection, where MobileNet outperformed ResNet50 with accuracies of 94.63% and 92.86%. Authors in ^[4] achieved a high accuracy of 98.75% on peach disease detection using a CNN model with the PlantVillage dataset. In contrast, our study employed a hybrid dataset of 36,600 images across nine classes of stone fruits and utilized several pre-trained models, with ResNet50 emerging as the best, achieving 93.11% accuracy. This demonstrates that larger and more diverse datasets, combined with robust models like ResNet50, significantly enhance performance in crop disease detection.

7. Conclusion and Future Scope

The experiments in this study are restricted by the use of free resources (Google Colab). Since Google Colab only provides the server for a short period of time, GPU is not available all the time. It takes about 8 to 9 hours to train a CNN model with a dataset size of 30,000-35,000 and 30-40 epochs if a GPU is not available. Therefore, the experiment related to customization of ResNet50 was not conducted in this study. An additional constraint of the study is that it relied on publicly accessible secondary data rather than primary data that was actually gathered from the field.

Since many writers have addressed the issue of data scarcity, as we indicated in the introduction section, it is important to support data gathering efforts in this area to make it simpler for researchers to conduct experiments. This study shows that, ResNet50 beats other three CNN models–MobileNetV2, DenseNet201, and InceptionV3. ResNet50 had the highest accuracy thus, we can modify the ResNet50 model to further improve its accuracy; this will result in an improved ResNet50 model. The experiment concludes that ResNet50 can manage the vanishing gradient problem, is low-power and parameterized to suit resource limits, and is faster than MobileNetV2, DenseNet201, and InceptionV3.

Our future goal is to enhance the accuracy rate ahead by creating a new deep-learning architecture that draws inspiration from ResNet50. Additionally, we aim to develop a localization method that applies image segmentation techniques to detect the diseased area on a leaf.