Edge Architecture for High-Accuracy Disease Identification in Apple Plants Using Transfer Learning Approach

Over the years, both India and the United States have witnessed substantial growth in apple production, contributing significantly to their agricultural sectors. India's apple production surged from 185,000 tonnes in 1961 to an impressive 2,276,000 tonnes in 2021, while the United States saw its production increase to 4,467,206 tonnes in the same period [1]. This growth has been supported by technological innovations, government initiatives, and advancements in agricultural practices. Despite these achievements, both countries face a persistent challenge that threatens the sustainability of this success: the accurate and timely identification of crop diseases.

The current research introduces an innovative solution to this problem through the implementation of smart agricultural methods, utilizing cutting-edge technologies such as Artificial Intelligence (AI), the Internet of Things (IoT), and machine learning algorithms [2]. These technologies enable the automated detection and identification of diseases in agricultural settings, employing various sensors, drones, and imaging techniques to analyze plant health [3, 4]. Such advancements offer a significant improvement over traditional methods, which largely rely on manual inspection and are subject to human error and biases [5, 6].

Despite the advent of smart agricultural methods, the challenge remains in achieving high accuracy in disease identification, especially in remote and rural areas where access to cloud-based systems is limited. The reliance on cloud architecture for processing and analyzing data can introduce latency, require stable internet connectivity, and raise concerns about data security and privacy [7, 8]. Additionally, the cost of implementing cloud-based solutions can be prohibitive for small-scale farmers [9, 10].

To address these challenges, this paper proposes a novel edge architecture-based system for high-accuracy disease identification in apple plants using a transfer learning approach. Edge architecture enables real-time data processing and analysis at the source – directly on edge devices or within a local network. This minimizes latency, reduces dependency on internet connectivity, and addresses privacy and security concerns by keeping data localized [11, 12]. AI algorithms, deployed at the edge, can analyze data in real-time, detecting early signs of diseases or stress in plants, thus enabling timely intervention to prevent the spread of diseases and minimize yield losses [13, 14].

This work seeks to fill the gap in current research by providing an efficient, cost-effective, and scalable solution for disease identification that can operate independently of cloud-based systems. The edge architecture-based system represents a significant advancement over existing methods by offering a decentralized approach to disease detection, which is particularly advantageous for applications in remote or resource-constrained environments. By leveraging the power of AI and transfer learning, our system not only enhances the accuracy of disease identification but also contributes to sustainable agricultural practices, potentially transforming the management and outcome of crop diseases in apple production.

3.1 Dataset

This investigation leveraged a publicly accessible dataset, which is a portion of the one assembled by Plant Village [34, 35]. The comprehensive dataset encompasses 9,714 high-resolution images showcasing a variety of diseases found on apple leaves. These images were meticulously obtained under diverse conditions, incorporating variations in illumination, angles, and backgrounds, to guarantee an all-inclusive representation of potential scenarios. The dataset employed for this study includes 9,714 images of apple foliage, proportionally allocated across four categories: cedar apple rust, multiple diseases, healthy leaves, and apple scab. The primary purpose of this study is to precisely categorize apple foliage into these four distinct classifications [36].

Conventional Convolutional Neural Networks (CNNs) used for classification purposes are typically seen as an amalgamation of two primary elements: a block for feature extraction and a classifier segment. The role of the feature extraction block is to manipulate the input image via a sequence of convolutions, pooling processes, and linear activation functions. This results in the creation of a unique feature map at every step of the operation. Suppose we visualize a theoretical situation where $G_a$ symbolizes the feature mapping of the a^th layer within the CNN architecture. These features can be derived using the Eq. (1):

$G_a=\lambda\left(G_{a-1} R_a+s_a\right)$        (1)

In this equation, in the given network architecture, we utilize G_a to delineate the characteristic mapping of the present layer. Meanwhile, G_a-1 is employed to articulate the attributes from the convolutional operations of the preceding layer. The layer-specific weight parameters are symbolized by R_a, while s_a is indicative of the offset vector associated with the a^th layer. Furthermore, $\lambda$ embodies a function that draws its inspiration from the Rectified Linear Unit (ReLU) mechanism. Further pooling layers play a crucial role in reducing spatial dimensions, thereby managing complexity and addressing the overfitting problem in neural networks. The pooling layer's $p^{\mathrm{th}}$ feature as represented in Eq. (2) is obtained as an outcome from the $q^{\mathrm{th}}$ localized receptive field, employing a down-sampling function.

$V_q^p=\operatorname{down}\left(V_q^{p-1}, p s\right)$          (2)

This process involves utilizing the previous layer's feature vector, $V_q{ }^{p-1}$, and a specified pooling size, ps. By minimizing spatial dimensions, pooling layers help in creating more efficient and generalized representations of the input data. The structure of a neural network encompasses a series of densely interconnected layers that succeed the convolutional and pooling layers. The chief function of these densely interconnected layers is to leverage the features that have been isolated for categorizing the images. The 'Softmax' function is then utilized to scrutinize the inferences drawn from the previous layers, which are founded on the isolated features. The Softmax function is represented by Eq. (3), where '$\mathrm{M}$' denotes the dimension of the vector 'd'.

${Softmax}(d)=\frac{e^d}{\sum_{m=1}^M e^d}$        (3)

1.png

Figure 1. AI model building steps involved

The ultimate outcome of this block is a condensed feature vector, representing the input image with only the most significant semantic features, effectively compressing the information. The compact feature array is subsequently directed to the categorization unit, an integral part of the structure that consists of an interconnected neural network system [37-39]. The purpose of this categorization unit is to critically assess the feature array, and then generate a distinct array filled with likelihood values. These values represent the probability of the initial input image being associated with each individual category. The procedure involved in constructing an efficient system for distinguishing between healthy leaves and those affected by disease is depicted in Figure 1.

MobileNet_V2 was designed primarily for the task of image classification in the field of computer vision. It is a deep convolutional neural network architecture that focuses on efficient and lightweight models suitable for running on mobile and embedded devices with limited computational resources. The main goal behind MobileNet_V2 was to provide a network that could achieve high accuracy on image classification tasks while being efficient in terms of model size and computational cost. This makes it particularly well-suited for real-time applications on mobile devices, such as object recognition in smartphone cameras or other scenarios where computational resources are constrained.

In the concluding stages of the Convolutional Neural Network (CNN) structure, the categorization component processes the generated feature map via a fully connected layer that has outputs corresponding to the total number of classes [40]. This is accomplished by employing the normalized exponential function, or softmax, as an activation function, thereby yielding the final predictions. In this particular implementation, a convolution with a kernel size of (1, 1) was used. This approach effectively condenses all the RGB channels of the input images into three channels for the network's operations. It's crucial to note that this reduction isn't determined by choosing important bands; the resulting three channels don't represent specific bands, but rather, they are a projection of the information in the original bands. This is akin to using Principal Component Analysis for band analysis and selecting the three most significant components. The training process of this CNN involves learning the necessary transformation to achieve this reduction in channels [41-43]. In a similar vein, the classification section needed a revamp to cater to a quad classification problem, differentiating between healthy and infected samples. Consequently, the last activation map underwent linear compression, reducing it from 1024 to 64 channels, and was then converted into a one-dimensional representation by flattening it to obtain the feature vectors. The primary elements of the MobileNet feature extraction mechanism stay intact, while the remaining components of the application are customized to tackle our unique challenge. This adjustment ensures the solution is designed to target our specific problem, while maintaining the efficiency of the original feature identification structure inherent to MobileNet.

In essence, each RGB image showcasing apple foliage was initially adjusted to a resolution of 256 × 256 pixels. It was then converted into a tri-channel image using a linear transformation applied to each individual pixel, similar to the effect of a 1 × 1 kernel convolution, yielding three separate channels instead of bands. These channels were then fed into the pre-trained MobileNet_V2 model, which in turn generated an activation map of the dimensions 8 × 8 × 1024. This map was further condensed from 1024 channels down to a mere 64, a process accomplished via another convolution, this time utilizing a 1 × 1 kernel size. Ultimately, this final map was simplified into a one-dimensional form, resulting in a feature vector that was directly linked to the original image.

3.2 Methodology

This research employs a systematic approach to modifying the MobileNet_V2 architecture for the identification of diseases in apple leaves, focusing on hyperparameter tuning and architectural refinement to enhance performance on the current dataset. The methodology is structured into six key components:

Dataset Preparation: We curated a comprehensive dataset of apple leaf images, annotated with disease classifications. This dataset includes images of healthy leaves and those affected by common diseases. Images were pre-processed for normalization, augmentation, and splitting into training, validation, and testing sets to ensure model robustness and generalizability.

Model Selection: The MobileNet_V2 architecture was chosen as the base model due to its efficiency and effectiveness in image classification tasks. Its lightweight nature makes it suitable for edge computing applications, where computational resources are limited.

Hyperparameter Tuning: Extensive experiments were conducted to fine-tune the hyperparameters of the MobileNet_V2 model, including learning rate, batch size, and the number of epochs. The objective was to find the optimal settings that maximize accuracy while minimizing overfitting.

Architectural Modifications: We introduced specific modifications to the architecture to better cater to the nuances of apple leaf disease identification. This included adjusting the depth multiplier, layer freezing during transfer learning, and incorporating custom layers to enhance feature extraction capabilities specific to leaf disease patterns.

Training and Validation: The modified MobileNet_V2 model was trained on the prepared dataset, utilizing a combination of real-time data augmentation to enhance diversity and robustness. Validation was performed iteratively throughout the training process to monitor performance and prevent overfitting.

Performance Evaluation: The model's accuracy, precision, recall, and F1 score were evaluated against a held-out test set. Comparative analysis was conducted against baseline models and existing approaches to demonstrate the improvements achieved through the proposed modifications.

By meticulously adjusting the MobileNet_V2 architecture and optimizing it for the specific challenge of apple leaf disease identification, this study aims to set a new benchmark for accuracy and efficiency in agricultural disease detection applications.

3.3 Experimental validation

To ensure the robustness and effectiveness of our modified MobileNet_V2 architecture for identifying diseases in apple leaves, we meticulously designed an experimental validation framework. This framework encompasses the experimental setup, detailed training process, evaluation metrics, and a comprehensive validation strategy aimed at preventing overfitting while assuring model performance on unseen data.

Our experiments were conducted using a custom dataset comprising high-resolution images of apple leaves categorized into healthy and various disease states. The dataset was augmented to enhance model robustness against variations in lighting, orientation, and background. The modified MobileNet_V2 model was implemented in TensorFlow 2.0, leveraging a CUDA-enabled GPU for efficient training.

The model underwent a two-phase training process. Initially, the model's base layers, pre-trained on ImageNet, were frozen to transfer learned features to our task. Subsequently, the entire model was fine-tuned on our apple leaf dataset with a reduced learning rate to adapt the high-level features specifically for our disease identification task. Training was performed over 50 epochs with a batch size of 32, using the Adam optimizer for its adaptive learning rate capabilities. Data augmentation techniques such as rotation, zoom, and horizontal flipping were applied to increase the diversity of the training data, simulating a wide range of real-world conditions.

Evaluation Metrics: We evaluated the model's performance using accuracy, precision, recall, and the F1 score as primary metrics. Accuracy measures the overall correctness of the model, precision evaluates the model's ability to identify diseased leaves correctly, recall assesses the model's capability to detect all actual disease cases, and the F1 score provides a balance between precision and recall, offering a holistic view of the model's performance.

Adopting knowledge from a task and applying it to a similar, yet distinct task is a robust strategy in machine learning known as transfer learning. This technique proves exceptionally beneficial in the initialization of neural networks. It is often more effective to commence the training process with models that have been pre-trained on expansive and well-labeled datasets, such as the widely-used ImageNet, rather than initiating the process from the ground up with arbitrary weight initialization. In our exploration, we focus on the application of pre-existing models, honed on the comprehensive and illustrative ImageNet dataset. These models are then meticulously adapted to align with the specifications of the desired dataset. The key steps involved in transfer learning entail identifying the suitable base networks, constructing a new neural network, and fine-tuning the model to adapt it for the desired task at hand. By following these transfer learning methodologies, we can efficiently capitalize on the pre-existing knowledge captured by the base models and seamlessly integrate them into our specific task. By adopting transfer learning, we can significantly expedite the training process and achieve enhanced performance with less computational burden. The use of pre-trained models facilitates the extraction of meaningful features from the base dataset, which can be effectively leveraged to tackle the target dataset's intricacies.

Step 1: Identification of Base Networks

In the process of transfer learning, we identify the base networks that will serve as the foundation for our model. These base networks are allocated weights ($\mathrm{W} 1, \mathrm{~W} 2, \ldots, \mathrm{Wn}$) obtained from a pre-trained CNN model.

Step 2: Building the Neural Network

To create a new structure for our network, we can modify it by replacing, inserting, or deleting layers as needed. This step allows us to customize the architecture to suit the specific requirements of our task.

Step 3: Fine-tuning the Neural Network

With our own dataset d and its corresponding labels W, we proceed to fine-tune the newly constructed neural networks. The objective is to minimize the loss function (F) through this fine-tuning process, resulting in a more tailored model for our specific task as indicated in Eq. (4).

$\begin{aligned} F(R)=-\frac{1}{n} \sum_{d_x=1}^n & \sum_{y=1}^T\left[b_{x y} \log Q\left(d_x=y\right)\right] +\left(1-b_{x y} \log \left(1-Q\left(d_x=y\right)\right)\right)\end{aligned}$      (4)

In the framework of our research, we denote the weight attributed to the convolutional layers and fully-linked layers as 'R'. The parameter 'n' is indicative of the total count of examples utilized during the training process, with 'x' functioning as a marker for the particular examples in training. We also employ 'y' as a symbol to denote the index of classes. The likelihood of an input denoted by 'dx' falling into the class predicted as 'y' can be articulated as $\mathrm{Q}\left(\mathrm{d}_{\mathrm{x}}=\mathrm{y}\right)$. The process to gauge the precision of 'R' is executed by lessening the loss function (F) on the original set of data, as articulated in the succeeding Eq. (5).

$R_y=R_{y-1}-l\left(\frac{\tau F(R)}{\omega R}\right)$         (5)

2.png

Figure 2. Architecture of the proposed AI model

Figure 2 illustrates the architecture of the proposed AI model. This approach allows us to assess the model's performance and its ability to predict the correct class labels for the given inputs. By effectively adjusting the weights and optimizing the loss function, we can improve the model's accuracy and enhance its predictive capabilities. The learning rate denoted by "l" and the class index "y" play essential roles in the process. Consequently, we employ the MobileNet_V2 pre-trained model to conduct transfer learning on the images and train the newly acquired neural networks using our dataset. The method termed "fine-tuning" entails employing the pre-set weights from a pre-existing Convolutional Neural Network (CNN) as a launchpad for the training of a new CNN, followed by its gradual optimization. This procedure aids in the adaptability of the feature extractor to a novel domain. In the development of our model, we implemented this technique, leveraging a feature extractor from MobileNet_V2 that was pre-optimized on ImageNet. We then fine-tuned this using our own dataset of apple leaf images.

A pivotal component to achieving successful transfer learning is the careful training of a CNN, which ensures that the initial pre-set weights are not entirely overwritten during the re-optimization process. In our application, we elected to freeze the weights of the pre-optimized MOBILENET_V2, focusing solely on updating the weights of the newly introduced layers for a total duration of 20 epochs. An epoch, in this context, is defined as 12 iterations that employ Stochastic Gradient Descent (SGD), effectively covering our complete dataset once over. To reach this goal, we utilized the SGD variant known as the Adam optimizer, setting the learning rate at 0.0001 and processing 16 images per batch.

$Cross\,\,Entropy =-\sum_{i=1}^N \sum_{i=1}^M y_{i, j} \log \left(\widehat{y_{i, j}}\right)$          (6)

$Accuracy =\frac{1}{N} \sum_{i=1}^N \sum_{j=1}^M I\left(y_{i, j}=\hat{y}_{i, j}\right)$            (7)

We utilized the powerful capabilities of Python's TensorFlow 2.0 library to educate our Convolutional Neural Network (CNN). During this progression, we opted for categorical cross-entropy (as represented in Eq. (6)) as our loss function, courtesy of its beneficial mathematical characteristics for diminishing issues related to multi-class classification. For our main evaluative measure, we selected accuracy (as delineated in Eq. (7)), considering its excellent clarity when it comes to gauging the effectiveness of the model.

Validation Strategy: To prevent overfitting and ensure the model's generalization to unseen data, we employed a rigorous validation strategy. A hold-out validation set, comprising 20% of the dataset, was used to monitor the model's performance during training. Early stopping was implemented to halt training if the validation loss did not improve for ten consecutive epochs, preventing overfitting. Additionally, k-fold cross-validation was conducted post-training to further evaluate the model's robustness and reliability across different subsets of the data.