Comparison of Plant Leaf Classification Using Modified AlexNet and Support Vector Machine

Plants are the backbone for all living beings as they provide food. To have good quality and quantity of the food, we need to protect the plant from the disease. The priority action is to identify the species of the plant, and followed by identifying the diseases affecting them. Having a disease in plants is quite natural, so the detection of disease in plants is very important in the field of agriculture. The plant disease severity is an important parameter and thus can be used to predict yield. Appropriate care needs to be taken to avoid the major effects on plants as they, in turn, show their effect on the product quality, quantity, or productivity. The diseases in plants cause a heavy loss in production of yield and financial loss to the farmers.

Diseases in crops are mainly classified in two types viz. airborne and soil-borne. In the air-borne type, fungal diseases are very common. The symptoms of the affected plant are seen in certain part like leaves, stem, and fruit. In soil-borne diseases, the effect is seen majorly on the roots of the plant [1]. Various types of plant disease identification techniques are used. The very basic or traditional technique is the manual inspection of the plant by naked eyes. This process was required to be carried out by experts and requires continuous monitoring over a large area of a farm [2, 3]. This process is a time-consuming and expensive one. The faster and accurate identification of the severity of the disease will help to take preventive measures and reduce the yield losses. The recognition of plant disease using images that are captured from devices like mobile phone cameras or digital cameras proves to be a significant challenge.

To overcome this situation, we are looking for a quick, reliable, automated, cost-effective, and most importantly accurate method to detect disease in the plant. A technique that automatically detects the disease symptoms in the leaves of the plant is beneficial as it reduces a lot of manual work and saves time [4]. In most cases, whenever the plant disease is detected by the farmers, they just use chemical fertilizers to prevent the crop from further growth of the disease. This could lead to a hazardous effect on the crop as well as to the person coming in contact with that crop. Sometimes using some basic things like plucking the diseased leaf and burning it or using organic fertilizers also help in solving the problem. This all depends on how much percentage of the disease has affected the leaves of the crop.

Image processing is seen to be more successfully used in disease detection mechanisms. In recent times, various machine learning algorithms for plant disease classification for certain diseases and crops have shown promising results [5]. The computer-based image processing technology applied in agricultural engineering research has become common. The Decision Trees, K-means, k nearest neighbors, Support Vector Machines (SVMs), Artificial Neural Networks (ANNs), and Machine Learning (ML) are the basic techniques used for developing the model for classification purposes. The evolution of deep learning techniques has shown significantly better results as compared to the shallow ML algorithm. The deep learning network is being used in the detection of the disease on the plant leaf that will help in faster and accurate results. The botanists are now benefitted from the advances in science and technology with computer vision approaches in the plant identification task. Several perspectives have been proposed in the literature for the classification of plants. Here the work is proposing a convolutional neural network (CNN) model for the classification of the plant. 32 classes of 9 different plants are selected from the leaf database of the PlantVillage database for this purpose. The data is classified using a support vector machine and AlexNet with transfer learning.

The organization of the paper is as follows: Related work is in section 2, proposed work using support vector machine and deep learning is discussed in section 3. Results and discussion are in section 4, followed by the Conclusion in Section 5.

In this paper, pre-trained AlexNet with transfer learning is used for the classification of a plant leaf. Also, the performance of the network is compared with that of the traditional SVM. The proposed model for the classification of plants is as shown in Figure 1. The description of each block is explained in the further section from 3.1 to 3.5

1.png

Figure 1. Proposed model for classification of plant leaf

3.1 Healthy and diseased plant leaves dataset

The dataset of PlantVillage is taken for the classification of the plant leaf. The dataset consists of 9 different plants with healthy and disease cases. A total of 32 classes showing the variety of healthy and diseased plants from amongst the nine plants are there in the dataset. 100 images of each of the classes are selected to avoid the overfitting of data. The nine plants are apple, cherry, corn, grape, peach, pepper, potato, strawberry, and tomato. In these nine plants, one of the classes is healthy, and at least one is a diseased one.

3.2 Pre-processing of the data

For the smooth functioning of any algorithm and also to maintain uniformity in the analysis, it is necessary to follow the basic steps that are common throughout the analysis. Pre-processing is one of them. In the proposed work, SVM and AlexNet are used for classification purposes. For the AlexNet, the input requirement is that the size of the image should be $227 \times 227 \times 3 .$  The raw images chosen from the PlantVillage dataset are of size $256 \times 256 \times 3$. So, all input images are resized to the required format for AlexNet.

3.3 Generating training and testing dataset

The dataset is divided into two sets viz. training and testing dataset. In this work, an analysis of five different variations of the training dataset and the testing dataset is done. The network is trained with 10% training data, and the network is tested over the remaining data of 90%. Then the network is trained with 30% training data, and the network is tested over the remaining 70% of the data. The network is further trained with 50% training data, and the network is tested over the remaining data of 50%. Again, the network is trained with 70% of the training data, and the network is tested over the remaining data of 30%. And finally, the network is trained with 90% of training data, and the results are tested with the remaining data of 10%. These five combinations are used for training the network for the classification of the plant leaf. Table 2 shows the number of images in training data and testing data for different combinations of the total dataset, consisting of 3200 images.

Table 2. The number of images in the training and testing data set for different training data size

3.4 Modify and apply a deep learning algorithm

The proposed idea of classification of the plant leaf is first done using the proposed AlexNet with transfer learning, and the classification is done by the state-of-the-art technique SVM with linear kernel and then using the SVM with Radial basis function (RBF) kernel. Classification of these techniques is compared with the deep learning classifier network.

3.4.1 Support Vector Machine

SVM is a supervised learning classifier. A decision plane in the SVM classifier splits between a set of objects partaking different class members. A linear classifier separates the classes into respective groups with a line. The classification of the objects is based on distinguishing the classes by drawing a separating line known as a hyperplane. SVM minimizes the high leap of the generalization error and maximizes the boundary that is created between an unravelling hyperplane and training data. SVM solves the problem caused by the error due to local minima; overfitting etc. The training data in the input space is mapped by SVM into a high-dimensional feature space. In this paper, the SVM with linear kernel and SVM with RBF kernel is used. The linear decision boundary of the feature space created is determined by generating the hyperplane distinguishing the classes. The linear SVM model scales linearly as per the size considered for the training dataset. The RBF kernel of SVM is proposed to accelerate the training time of the soft margin in the support vector machine [17]. In the SVM with RBF kernel $\left\{x_{j}^{(i)}\right\}_{j=1,2, \ldots N_{i}} \subset R^{d}$  is the set of samples from the training data in class $i$, where $N_{i}$ is the number of training samples in class $i$, with $i=1,2, L \text { and } L$ is the number of classes. The RBF kernel is

$k\left(x, x^{\prime}, \sigma\right)=\exp \left(\frac{\left\|x-x^{\prime}\right\|^{2}}{2 \sigma^{2}}\right)$    (1)

where, $x, x^{\prime} \in R^{d}-\{0\}$ is the corresponding parameter.

3.4.2 Algorithm for classification of plant leaf using Support Vector Machine with a linear kernel

The Algorithm for the classification of the plant leaf of the dataset of nine plant leaves is shown below. Step 1 to step 7 are followed for all the algorithms used in the paper. All the algorithms are implemented in MATLAB 2017B. The input size is also taken commonly in all cases.

Step 1: Read the data of healthy and diseased plant leaf

Step 2: Pre-process the data: Resize the data

Step 3: Label the data with the plant name and the healthy or disease name. e.g., Apple Healthy, Apple Black Rot, etc. for all 32 classes

Step 4: Divide the data into a training dataset and a testing dataset. The data is split into a training dataset with the variation of size having 10%, 30%, 50%, 70%, and 90% of the total data, and the remaining is used as a testing dataset in each case.

Step 5: Train the network using the training dataset for the proposed model shown in Figure 2.

Step 6: Test the network using a testing dataset using the SVM with a linear kernel.

Step 7: Calculate the performance parameters. The classification of data into the healthy or diseased plant can be done based on the performance parameters like accuracy, confusion matrix, etc.

3.4.3 Algorithm for classification of plant leaf using Support Vector Machine with Radial Basis Function kernel

The Algorithm for the classification of the plant leaf of the dataset of nine plant leaves is shown below. Step 1 to step 7 are followed in this classification task.

Step 1: Read the data of healthy and diseased plant leaf

Step 2: Pre-process the data: Resize the data

Step 3: Label the data with the plant name and the healthy or disease name. e.g., Apple Healthy, Apple Black Rot, etc. for all 32 classes

Step 4: Divide the data into a training dataset and a testing dataset. The data is split into a training dataset with the variation of size having 10%, 30%, 50%, 70%, and 90% of the total data, and the remaining is used as a testing dataset in each case.

Step 5: Train the network using the training dataset for the proposed model shown in Figure 2.

Step 6: Test the network using a testing dataset using the SVM with radial basis function kernel.

Step 7: Calculate the performance parameters. The classification of data into the healthy or diseased plant can be done based on the performance parameters like accuracy, confusion matrix, etc.

2.png

Figure 2. Network layers of the proposed model based on AlexNet

3.4.4 Deep learning methods

Deep learning models are evolved from the basic neural networks. The difference between deep learning and conventional artificial neural networks is that it has multiple numbers of hidden layers between the input and output layers. The main layer in the convolutional deep neural network model is the convolutional layer. In these models, the raw input is fed to the network to fetch certain task-specific output at the final layer of the network. Deep learning has numerous applications in the field of classification of images or recognition of voice and pattern [18, 19] in a large database. The proposed model in the study [20] a deep learning model for the “ImageNet Large Scale Visual Recognition Challenge” (ILSVRC) dataset competition where his network AlexNet was able to classify the 1000 classes. There are various convolutional neural networks (CNN) models like AlexNet, GoogLeNet, ResNet, VGG16, VGG19, DenseNet, SqueezeNet, etc. The difference in these networks is the depth in the layers and the nonlinear functions that are used in them. Otherwise, the structure is the same consisting of four important layers viz. “convolution layer”, “max-pooling layer”, “fully connected layer”, and the “output layer”.

The AlexNet is a pre-trained network that can classify 1000 classes. For the proposed work to classify the 32 classes of the dataset, transfer learning is done at the last three layers of the network. The proposed model based on AlexNet is shown in Figure 2. The results of five combinations of training testing datasets are used for all the three networks mentioned above. The classification accuracy and confusion matrix, along with the simulation time, is noted in each of the cases. In the proposed work, the classification of the leaf of nine plant varieties with 32 different classes consisting of the healthy and diseased leaf is done. In this proposed network, transfer learning was used that helps in the classification at the output layer as per the required purpose showing 32 different classes. This network consists of an input layer followed by five convolutional layers than two fully connected layers and then transfers learning layers. The requirement for the input layer is the size of the image in a specific dimension. This criterion needs to be satisfied for any CNN model. The input is convolved with the weight vectors and depending on the padding and the stride used; the layer can take up the size that is the same as before or compact or expanded. Rectified Linear Unit (ReLU) is the most frequently used as an activation function in neural networks, especially in CNNs.

$f(x)=x^{+}=\max (0, x)$    (2)

where, $\mathcal{X}$ is the input to a neuron in the network.

The probability of a vanishing gradient can be reduced by ReLU. This also introduces sparsity to the model [21]. The ReLU nonlinearity layer of the first and second convolution layers is followed by a local normalization step before doing pooling. The necessary conditions for the computational purpose and the size of it can be reduced with the help of pooling. Max pooling shows better performance and high convergence. The data is down-sampled with the help of the max-pooling layer. The chances of overfitting are reduced by this. The dropout layer offers a remarkably effective regularization and computationally cheap method to reduce overfitting problem and improve the generalization error on CNN. The first two fully connected layers are densely connected and the last fully connected layer is modified as per our requirement to predict and classify 32 different classes.

Transfer learning is the process to modify the network at the final stage for the desired output. In the transfer learning, the last three layers of the network are replaced by the fully connected layer stating the number of classified outputs that is desired, followed by the softmax activation function layer, and last the classification output layer.

3.4.5 Algorithm for classification of plant leaf using the proposed model in Figure 2 with deep learning classifier

The Algorithm for the classification of the plant leaf of the dataset of nine plant leaves is shown below. Step 1 to step 7 are followed in this classification task.

Step 1: Read the data of healthy and diseased plant leaf

Step 2: Pre-process the data: Resize the data

Step 3: Label the data with the plant name and the healthy or disease name. e.g., Apple Healthy, Apple Black Rot, etc. for all 32 classes

Step 4: Divide the data into a training dataset and a testing dataset. The data is split into a training dataset with the variation of size having 10%, 30%, 50%, 70%, and 90% of the total data, and the remaining is used as a testing dataset in each case.

Step 5: Train the network using the training dataset for the proposed model shown in Figure 2.

Step 6: Test the network using a testing dataset using the deep learning classifier.

Step 7: Calculate the performance parameters. The classification of data into the healthy or diseased plant can be done based on the performance parameters like accuracy, confusion matrix, etc.

3.5 Classify the healthy and diseased plant leaf

The classification of the healthy and diseased plant leaf is evaluated with the performance parameters. The performance parameter used for the classifier here is accuracy, confusion matrix, and the time required for computation of the classification task.

3.5.1 Accuracy of the classified output

The accuracy for the predicted output from the classifier is calculated using Eq. (3).

$\text { Accuracy }=\frac{\text { Number of correctly classified classes }}{\text { Total input classes }}$     (3)

3.5.2 Confusion matrix

The performance of the classifier is measured with the help of the confusion matrix. The confusion matrix consists of classes that are correctly classified and the classes which are misclassified. The parameters in the confusion matrix are “true positive”, “true negative”, “false positive”, and “false negative” [22]. In this work, there are 32 classes, so the confusion matrix is a multiclass one with the size of $32 \times 32$.

3.5.3 Simulation time

Time is significant in any aspect. Here the time elapsed for training and testing of the classifier for each of the case of training data of 10%, 30%, 50%, 70%, and 90% for each of the classifier i.e., SVM with linear kernel and RBF kernel, and pre-trained AlexNet with transfer learning is measured. The time is measured in seconds.