Potato Leaf Disease Classification using Transfer Learning based Modified Xception Model

- Plant diseases are the essential thing which decreases the quantity as well quality in agricultural field. As a result, the identification and analysis of the diseases are important. The proper classification with least data in deep learning is the most challenging task. In addition, it is tough to label the data manually depending upon the selection criterion. Transfer learning algorithm helps in resolving this kind of problem by means of learning the previous task and then applying capabilities and knowledge to the new task. This work presents the convolution neural network-based model to predict and analysis the potato plant disease using plant village datasets with deep learning algorithms. Transfer learning with feature extraction model is employed to detect the potato plant disease. The results show that improved performance with an accuracy of 98.16%, precision of 98.18%, the recall value of 98.17% and the F1 score value of 98.169 %.


Introduction
In today's world, agriculture with modern technologies produce enough food to meet the people demand and economy of India also depends on its productivity [1,2]. However, security of food remains challenging owing to factors such as climatic changes, the decline in pollinators, plant diseases, and other factors. Potato is the fourth foremost used world's significant food crop after rice, wheat and corn according to the statistics of the survey done in agriculture department of USA [3]. Potato production exceeds 300 million metric tons across global and it contains a various compound including vitamins, minerals, proteins and carbohydrate which is the vital elements for assist and maintain the human health [4] in proper way. Phytophthorainfestans (late blight) and Alternariasolani (early blight) are the two common disease occurs generally in potato plants [5].Detection of plant diseases play vital role in agricultural field .Moreover, plant diseases are becoming big threat to food security and small holder farmers. In the developing nation, small holder farmers contribute more than 80 percent of agricultural production and reports 50 percent of crop yield loss due to diseases and pets [6].
Automated classification of plant diseases based on images is a complicate problem owing to inter class similarities of plant and extrinsic factors including variations in image background, illumination, color, pose and occlusion [7].An early detection of diseases allows taking preventive measures against huge damage, curtailing the production and economic losses to the farmers. Over the last decades, plant diseases is identified by specialist by their naked eyes, but this approach results infeasible due to unavailability of specialist at farm situated in the remote locations and too much of processing time. Hence, the introduction of computer vision, machine learning, image processing and deep learning techniques turn out to be an efficient way to detect the plant disease at early stage and continuous monitoring of plant health condition.
Deep Learning (DL) is sub-field of machine learning, which is a part of artificial intelligence. In recent years, DL has achieved significant success in various field including speech recognition, natural language processing, computer vision, image analysis and recognition. This paper mainly focuses on detection and recognition of diseases from leaves of the potato plants using deep learning. The potato leaf images including late blight, early blight and healthy plant images are taken from the plant village website which are trained and tested on Modified-Xception model for classification of diseased and healthy plant.
The structure of this paper is organized as follows: section 2 gives insight about the related literature work. Materials and methodology are discussed in section 3. Section 4 provides the performance evaluation of the proposed method. Section 5 deals the experimental results and its discussion on classifying the potato plants. At last, conclusions are drawn in the section 6.
Advances in Computing, Communication, Automation and Biomedical Technology

Related works
The literature related to detect and recognize the plant disease through plant image datasets with deep learning algorithm is discussed in this section. [8] proposed an automatic disease detection system based on web to identify the disease affected in pomegranate fruit. Bacterial blight is the major disease which decreases the fruit production and causes the loss for the farmers. The system proposed K-means algorithm for clustering based on extracted features including color, morphology, CCV and finally Support Vector Machine(SVM) is employed for the classification of disease infected fruits and healthy fruits with 82% of accuracy. [9] presented machine vision based agro-medical expert system in order to assist farmer for papaya cultivation and detection of diseases. The system employs k-means clustering algorithm to segment the disease affected spot and extract the required features from the image captured through mobile or handheld devices. Then the features are used by Support Vector Machine (SVM) and achieve more than 90% of accuracy in classification of papaya diseases [10]. In [11], the author used artificial neural network (ANN) and various image processing technique for early detection of plant diseases. The proposed approach extracts the features through Gobar filter and results recognition rate up to 91% using ANN classifier. The classifier based on ANN performs classification of different plant diseases and uses the color, texture features combination to recognize plant with different diseases.
The author in [12] proposed a transfer learning approach for classification of tomato crop disease using pre-trained deep learning architecture especially AlexNet and VGG16.The architectures are trained for classifying seven different categories (six disease classes and a single healthy class) of tomato leaf images, consist of 13,262 segmented leaf images obtained from the public dataset website of PlantVillage.The architectures AlexNet and VGG16 achieve 97.49% and 97.23% of accuracy respectively in classification of diseases. The experimental results shows that VGG16 outperforms compared with AlexNet.
[13] developed a model based on convolution neural network to perform identification and diagnosis of plant diseases through deep learning methodologies using healthy and diseased plant images. The models were trained with 87,848 images, including 58 distinct classes of plant and disease. The VGG convolution neural network outperforms among all the models with success rate of 99.53 % in classification of plant diseases. [14] proposed a method for detection of rice disease using techniques, deep convolution neural networks (CNNs). The CNNs are trained on dataset, which consist of 500 rice leaf images including diseased, healthy and stems captured from the rice field. The model results 95.48% of accuracy in recognizing 10 common diseases of rice through 10-fold cross validation. Compared with other conventional methods, the proposed model outperforms in training the CNNs parameters, convergence speed and recognition rate.

Image Dataset
The datasets related to this study are collected from the plant village -an open access repository. This repository consists of 14 varieties of plants with 54,306 leaf images of healthy and diseased plant categories. In order to authenticate the proposed research work, the datasets of potato leaves are available in the plant village database. This database comprises of 3 different kinds of classes in which 2 classes are diseased and 1 class comes under healthy category. The disease occurs in the potato plants and its sample images with quantity of images in each class are tabulated.

Data Augmentation
Data augmentation method is applicable for increasing the data size utilized to train a model. In order to obtain the consistent identification, the deep learning algorithms frequently needs an enormous number of datasets for training that is not possible for all the times. As a result, the available datasets are augmented to create better comprehensive datasets to be needed for further processing. An over-fitting takes place in CNN algorithm when the model attains the best fit through learning the patterns available in the training datasets and fall short to recognize when the data is not available in the training datasets. From the Table 1, the images for healthy plant category is 152 and the other two disease affected images are in 1000 in quantity. In order to balance the images in all categories, data augmentation techniques like flipping and rotations are utilized. The samples after performing the data augmentation techniques are depicted in Figure 1.

Transfer Learning
The CNN models need a numerous dataset for training from the scratch data that is the most difficult task for the researchers who are working in this field. As a result, machine learning technique namely transfer learning is utilized in most of the cases. Transfer learning (TL) applies the knowledge of the CNN model trained on the large dataset to the similar smaller dataset. TL is also a very famous method used in pre-trained CNN models trained with ImageNet dataset. TL is employed for two categories namely feature extraction and fine tuning. In feature extraction process, the output layer of the pretrained model is truncated and lower layers act as the feature extractor tool. In fine-tuning, the weights of the lower layers are frozen and newly added top layers are updated through retraining with the new images.

Modified-Xception model
The Xception model is trained on ImageNet dataset that comprises of 1.2 million image datasets of thousand varieties of group. The proposed model is the modified version of the original xception model. In the Modified-Xception model the dense output layer is replaced with the new softmax layer which is depicted in the Figure 2 The Modified-Xception model performs of three main tasks such as feature extraction, fine-tuning and disease classification.

Feature extraction
The feature extraction is the most important component of CNN which involves convolution layers, pooling layers and activation functions. The activation function used in this study is Rectified Linear Unit (ReLU) which is used for activating the neurons in each convolution layer [15][16].

Convolution layer
Convolution layer is the primary layer in the CNN architecture which has filters for extracting the image features. The feature map FM from each convolution layer is achieved through sliding the filters on the input images. The convolution operations take place between input image pixels and filter pixels which produces the feature map. The feature map from each convolution layer is obtained using the Equation (2) Global Average Pooling (GAP) layer GAP obtain the feature map for every object used in the classification task from the final convolution layer. Then, GAP generates the feature vector FVGAP and forward to the softmax layer for classification. The feature vector FV GAP is obtained using the Equation (3).
Disease Classification Fine tuning Fine-tuning happens in the top most layer of CNN model which improves the performance of the model. The weights of the top layers are updated by means of retraining the layers with target datasets. The output layer weights are updated with retrained potato disease datasets. The fine tuning done in this work follows the Equation (4).

FinT = ft X FV GAP
(4) Softmax Classifier Softmax classifier uses output obtained from APL layer as input for its operation. From the input data, it estimates the probability by means of Equation (5). for each individual class. The resultant class is the class in which the maximum probability occurs. for i=1………….k (5)

Performance Evaluation Metrics
The predictive strength of the model is evaluated using the performance evaluation metrics viz precision, accuracy, F1 score and recall. The accuracy can be calculated using Equation 6 which gives the value by number of images identified correctly among the total number of predictions done.
The precision measurement is done using Equation (7) which is the ratio of exactly identified positive results (TP -True positive) to the total quantity of positive results identified by the proposed model.

FP TP
TP ecision + = Pr (7) where FP (False positive) is the number of wrongly identified positive results. The precision value becomes less value when FP value increases. The range of precision metric is in between 0 to 1.
Recall metric is measured with correctly identified true positive predictions (TP) divided by overall positive samples (TP+FN) shown in Equation (8).
FN TP TP + = Recall (8) where, FN denotes false negative The F1 score is obtained by Equation (9)

Experimental Results and discussions
In this research work, the model is trained with maximum number of 10 epochs. The choice of epoch is used for analyzing the performance of the model. Figure 3 presents the results of accuracy of training and validation datasets. Form the results, the accuracy value of the training datasets raises abruptly up to 5 epochs then it gradually increases for the upcoming values of epochs. In the same way, validation dataset results are also following the same way as the training dataset. In addition, training and validation loss for each individual epoch via model training is illustrated in Figure 4 which shows that the proposed model loss pertaining to training and validation datasets are higher in initial stages after that it slowly decreased.    The performance evaluation of the proposed work is done by calculating the accuracy of each test dataset. The accuracy performance is obtained using the values of Top 1 and Top 2 of the model is 98.16% and 100 %. The results of the performance evaluation metrics are listed in Table 3.

. Conclusion
Recently, agricultural field becomes automated for cultivation. The disease prediction in plants is needed for increasing the production yield. In this work, efficient usage of transfer learning with deep convolution neural network algorithm to classify the disease occurs in the potato leaves with least datasets. The proposed work depends on transfer learning with feature extraction of a pre-trained model done in ImageNet. The performance evaluation from the simulation reveals that the accuracy of the proposed work is to be 98.16 %. The outcome of this work shows that employing CNN model with transfer learning method can be used for plant disease classification with low accuracy performance scenario. This work can be extended to detect the disease in variety of crops such as cottons, crop millet etc.