Leveraging Inception V3 for Precise Early and Late Blight Disease Classification in Potato Crops

One of the important causes restricting agricultural produce and adversely affecting the economy of the nation and the income of farmers is the presence of pests and crop diseases. Early and non-destructive identification of pests and crop diseases is the most vital factor in the development of precision agriculture [1-4].

Potatoes (Solanum tuberosum) are a variety of edible plant tubers that grow in the roots of the plant. It is one of the most important food crops and has a high consumption rate across the world. Potatoes are highly vulnerable to infections of fungal and bacterial types; thus, disease-free, healthy potato plants have to be planted with much care in the growing areas. The wide range of diseases causes difficulty in the prediction, diagnosis, and control measures of their spread and occurrence. Devastating consequences and severe losses have come up as a limitation in potato cultivation [5].

Apart from being majorly used in cuisine, potatoes also contribute to the economic growth of many nations, including New Zealand, the US, and the UK; thus, millions of hectares of land are dedicated around the globe to potato cultivation. Potatoes are also a rich source of antioxidants, fiber, carbohydrates, minerals like potassium, magnesium, etc., and various vitamins, including vitamins B6 and C. They are a staple and versatile food in many regions and are found to be naturally gluten-free. They may help in controlling blood sugar levels and also improve the digestive health of humans [6, 7]. These medicinal benefits of potatoes add extra importance to the cultivation of potatoes. Many plants have medicinal benefits in their shoot parts, like lettuce, barley, etc. [8, 9]. The growth of medicinal herbaceous plants can also be enhanced by cultivating them in smart precision agricultural setups using IoTs or green IoTs [10]. Great losses occur in potato cultivation due to several factors, which in turn disturb the economies of productive nations. The major causes of potato disease are pests and plant diseases. Early detection of the diseased leaves of potato plants may go a long way toward saving the crop. This is becoming a major challenge due to the wide variety of crop species that are affected in different ways when subjected to different environmental conditions, and it leads to a delay in the detection of leaf diseases. Several machine learning and deep learning algorithms have been developed to help in the early detection of potato leaf diseases [11]. A major issue found in the existing systems is the efficiency obtained by implementing the algorithms.

Table 1. Classification of fungal plant diseases

Image	Disease Name	Causes and Symptoms	Climatic Conditions
11111.png	Early Blight	It is caused by a fungus named Alternaria solani. Infected leaves develop dark brown patches of 3-4 mm in diameter.	Early blight is found in cool, moist climatic conditions. Temperatures ranging within 75°-84°F and prolonged wetness of leaves due to frequent rainfall or sprinkler irrigation, together cause early blight infections.
2222.png	Late Blight	It is caused by Phytophthora infestans, an oomycete pathogen. Late blight is characterized by a circular brown infection, surrounded by pale greenish-yellow tissue.	Late blight is found when temperatures are mild and humid for prolonged periods. Temperatures ranging within 60°-78°F and humidity above 90%, together cause early blight infections.

Table 1 describes the two types of potato leaf diseases: early blight and late blight. These are fungal diseases that cause increased damage to plants in climatic conditions that enhance the growth of fungal infection.

A wide range of fungal and bacterial diseases affect the potatoes on the foliage and later spread throughout the entire plant. This causes serious damage to the block as a whole. Many of the diseases destroy the tubers first and then affect the leaves, causing the plant to wilt. Irrespective of the initial occurrence of the disease, the control action against fungal and bacterial diseases in potato plants must not be delayed [12]. Smart precision agriculture is useful in promoting the development of environmentally sustainable farming methods, ensuring food safety, enhancing economic growth, and protecting human health.

The proposed potato leaf disease detection model was trained in 1937 and validated on 215 images from a dataset of 2152 images of potato leaves. This technique proves to achieve 98.60% accuracy on this dataset. Finally, the performance of the proposed technique is compared with the existing state-of-the-art models and found to be better in terms of accuracy. At first, the Exploratory Data Analysis (EDA) is done, which helps us analyze the entire dataset and summarize its main characteristics, like class distribution, size distribution, and so on. It consists of brief analyses of the dataset considered so that the modeling process gets a direction and addresses preliminary queries. It may comprise the distribution of variables and verifying relevant patterns for various classes. Visual methods are often used to display the results of this analysis. Image pre-processing is carried out in the next step, where the features of the raw data are improved by suppressing the unnecessary distortions and making the data more suitable for the model by resizing and enhancing the important features and improving the performance. This holds for image classification as well, from which meaningful information can be extracted.

The authors [5] have proposed an improved deep learning algorithm that classifies the potato leaf images into five classes with 97.2% accuracy. The authors [12] used CNN to classify five types of potato leaf diseases on a dataset of 5000 images and obtained improved accuracy over R-CNN, AlexNet, Transfer Learning, VGG, and GoogLeNet. There are numerous challenges involved in the process of crop disease detection, namely: the evolution of new pathogens, generalization of the used model, technological dependency, collection of high-quality data, etc. [6-13].

In this research work, the authors have used a convolutional neural network (CNN), namely, Inception V3. Inception V3 is a deep learning (DL)-based CNN model mainly used for image classification. It is the third version of Google's Inception CNN, which is advanced and optimized to assist in image analysis and object detection. Along with restraining the count of parameters from increasing beyond a limit, the design of Inception V3 aims at allowing deeper networks. Some of the advantages of using the Inception V3 model as compared to its contemporaries like GoogLeNet, BN-Inception, PReLU, and VGG Net can be listed as [13, 14]:

• It shows better efficiency than the Inception V1 and V2 models.
• It shows a lower error rate while maintaining the same speed as compared with the previous models.
• It is a more advanced version and has a deeper network as compared to the previous versions.
• It is comparatively less expensive.

1.png

Figure 1. Depiction of the working of a pre-trained model when used to classify a new image dataset

The motivation for using transfer learning techniques for crop disease detection lies in the fact that a deep learning model takes a substantial amount of time to be trained and also requires a lot of samples to be trained with millions of parameters, whereas in transfer learning models, these problems are overcome. The flaws of deep learning models are overcome by transfer learning (pre-trained) models. In pre-trained models, the training time is reduced, which also ensures better accuracy. This research gap in classification problems is overcome using pre-trained transfer learning models. Inception V3, ResNet 50, VGG 19, and AlexNet are some of the widely used transfer learning models. With the help of transfer learning techniques, knowledge is transferred from a pre-trained model to another model that will perform classification. This reduces the training time, so the training time is saved. Figure 1 represents the workings of the transfer learning model to train a new model to classify a new image dataset.

The key contributions of this research are pointed out as follows:

(1) Collecting the image dataset of potato leaves and preprocessing it.

(2) Augmenting the dataset and splitting it into training and validation datasets.

(3) Training the Inception V3 CNN model to efficiently recognize and classify two types of diseased potato leaves and healthy leaves.

(4) Classifying the dataset with enhanced efficiency and reduced time.

(5) Comparison with the existing state of the art.

This approach significantly reduces the error rate and gives an accuracy of over 98%.

The rest of the article is organized as follows: an extensive background study in this regard has been done in Section 2, materials and methods are discussed in Section 3, results and discussion are described in Section 4, Section 5 depicts the comparison of the obtained results with the existing works, and Section 6 presents the conclusion and future work of the research work.

The materials required to carry out this research are a dataset of potato plant leaf images, duly split into training and validation datasets in the ratio of 80:20, and classified into three classes.

Data imbalance, a commonly faced challenge of uneven data representation, poses a key challenge to the biases of the learning algorithm. The algorithm will be biased toward the majority of the class of data, even though the majority of the data is not important. This challenge can be overcome using resampling.

Resampling is an effort to level up the unbalanced data. There are two basic ways of resampling:

• Over-sampling à it is the method of adding copies of items in the under-represented class. Figure 2 depicts the process of over-sampling.
• Under-sampling à this method comprises reducing the number of items from the over-represented class. Figure 3 depicts the process of under-sampling.

2.png

Figure 2. Process of oversampling

3.png

Figure 3. Process of under-sampling

The results obtained from implementing the Inception V3 model are discussed in this section.

5.1 Compiling and training Inception V3

Figure 9 is an illustration of the Inception V3 module. Each of the previous layers is passed through various convolution layers as well as one max pooling layer. After processing each of these layers, filter concatenation is performed.

The next step is to compile the model, optimize it using an Adam optimizer, and calculate the loss using a loss function. The metrics refer to the metric that has to be calculated, which is ‘accuracy’ in this case. After the compilation is done, the model has to be fit to the training data, or, in other words, the model has to be trained. The authors of this study have trained it in batch sizes of 128 with 10 epochs, using 80% of the data as the training dataset and the remaining 20% of the data as the validation dataset. A sequential model is created in TensorFlow Keras that comprises various layers of data augmentation. The layers are defined in Table 2.

9.png

Figure 9. The Inception V3 module

Table 2. Data augmentation layers

Table 3. Detailed specification of various training parameters for the CNN model

Table 3 comprises details about the various parameters used in the model, like activation functions, epochs, batch size, etc., used in the process of classification.

5.1.1 Accuracy

The evaluation measure used to calculate the classification accuracy is defined in Eq. (2). The number of right predictions divided by the total number of predictions gives the accuracy value.

Accuracy $=\frac{\text { Number of Correct Predictions }}{\text { Total Number of Predictions }}$                           (2）

The model, when trained with 10 epochs with a batch size of 128, gives an accuracy of 98.60% with a validation loss of only 3.63%. Figure 10 shows the accuracies obtained in each epoch.

Figure 11 shows the classification results of the potato leaf dataset. It can be seen that the maximum predictions are correctly made by the Inception V3 model that is used in this study.

10.png

Figure 10. Accuracies obtained in each epoch

11.png

Figure 11. Prediction results of potato leaf dataset

5.2 Plotting accuracy of Inception V3

Evaluate’ is a method to evaluate a CNN. The test data are taken as the parameters in this method. Priorly, the 'matplotlib' library and 'show' methods were used to plot the data on the console. Considerably good results were obtained for a specific type of problem using CNN. Figure 12 (a) shows the plot of training and validation accuracy, and Figure 12 (b) shows the plot of training and validation loss concerning different epochs.

12a.png

(a) Plot of training and validation accuracy

12b.png

(b) Plot of training and validation loss

Figure 12. Plot of accuracies and losses

5.3 Calculation of other quantitative metrics

Apart from accuracy, there are various other performance metrics like precision, recall, F1-score, mean square error (MSE), root mean square error (RMSE), etc. The authors have calculated the precision, recall, and F1-score in their experiment to better understand the performance of the model.

5.3.1 Precision

Precision is a performance metric used to measure the accuracy of the positive predictions of a model by calculating the correctly predicted positive observations above the total positive observations. The formula to calculate the precision is depicted in Eq. (3).

Precision $=\frac{\text { Number of Correctly Predicted Positive Instances }}{\text { Total Number of Positive Predictions }}$                                （3）

5.3.2 Recall

Recall is the performance metric used in deep learning classification algorithms, which is used to measure the true positive results out of all the positive cases in the dataset. The Eq. (4) depicts the formula for recall. Recall is alternatively called sensitivity.

Recall $=\frac{\text { Number of Correctly Predicted Positive Instances }}{\text { Total Number of Positive Instances in the Dataset }}$                         （4）

5.3.3 F1-score

F1-score is the combination of precision and recall, primarily used to strike a balance between these two metrics in cases of uneven class distribution where there is a need to minimize false positive and false negative cases. The formula to calculate F1-score is given in Eq. (5).

F1 - score $=2 * \frac{\text { Precision }- \text { Recall }}{\text { Precision }+ \text { Recall }}$                    （5）

F1-score plays a vital role in assessing the effectiveness of any classification model where there is a lack of clear preference between precision and recall. A tabular representation of all the performance metrics is shown in Table 4.

Table 4. Performance metrics

The Inception V3 model is used in this experiment because it has several significant advantages over other CNN models in various computer vision tasks. It used multiple parallel filters, which reduced computational costs considerably while conserving resources [6-9]. It is used as a pre-trained transfer learning model that reduces training time while attaining higher accuracy [9, 15].

Integration of AI techniques in agricultural practices to mitigate crop yield loss due to infestations from early-blight and late-blight diseases holds significant and promising importance in potato cultivation. A rapid detection of early and late blight diseases in potatoes at an early stage can be done with the help of AI-powered systems using image recognition and the analysis of sensor data in smart precision farming setups. Integration of AI also helps in the accurate and precise spraying of pesticides, thus ensuring their optimal use in the farming process. Crop management or recommendation systems can also be developed, which may help in accurate predictive yield analysis and crop planting and irrigation recommendations, ensuring another dimension of efficiency in farming. Such kinds of technologies provide authentic and analytical insights to identify patterns and correlate them.

This work was supported by the Deanship of Scientific Research, Vice Presidency for Graduate Studies and Scientific Research, King Faisal University, Saudi Arabia, under Project GRANT5,435.