Deep Transfer Learning Based Rice Plant Disease Detection Model

In agriculture, plant diseases are mainly accountable for reduction in productivity and leads to huge economic loss. Rice is the essential food crop in Asian countries and it gets easily affected by different kinds of diseases. Because of the advent of computer vision and deep learning (DL) techniques, the rice plant diseases can be detected and reduce the burden of the farmers to save the crops. To achieve this, a new DL based rice plant disease diagnosis is developed using Densely Convolution Neural Network (DenseNet) with multilayer perceptron (MLP), called DenseNet169-MLP. The proposed model aims to classify the rice plant disease into three classes namely Bacterial Leaf Blight, Brown Spot, and Leaf Smut. Initially, preprocessing takes place in three levels namely channel separation, grayscale conversion, and noise removal using median filtering (MF). Then, the fuzzy c-means (FCM) based segmentation process identifies the diseased portion in the rice plant image. The pretrained DenseNet169 technique is used as a feature extractor and the final layer is replaced by the MLP to perform rice plant disease classification. The effectiveness of the proposed model has been validated against benchmark dataset and the simulation outcome is examined under diverse measures. The obtained results defined the superior results of the DenseNet169-MLP model over the recently presented methods with the maximum accuracy of 97.68%.

techniques for improvising the entire crop productivity. On the other hand, rice is considered as a major food source provider of protein and energy to atleast 50% of the global population. Besides, the utilization of rice and demand gets increased with a rise in population. For fulfilling the rising food requirements, rice productivity should be raised to greater than 40% by 2030. However, rice plant diseases have resulted in a significant loss in productivity. Presently, the utilization of pesticide and blast-resistant approaches are the major ways available to combat the disease. The abundant utilization of pesticide not only increase the production cost, but it also affects the environment in a significant way.
At present times, computer-aided diagnosis (CAD) models are available to monitor crop diseases and pests using plant images. An automated rice disease diagnosis model can provide details to prevent and control the rice diseases for reducing the financial loss, decrease the pesticide residue, and raise the quality and number of crops. To attain such a model, researches are urged to develop effectual image processing approaches to detect plant diseases. The image processing steps involved in rice disease detection are preprocessing, segmentation, feature extraction, and classification processes. These processes are carried out only on the exterior appearance of the affected plants [2]. Generally, leaves are considered as the major source for plant disease detection. The plant diseases normally vary in size, shape, and color and each disease possesses some unique characteristics. Few diseases appear in yellow or brown in color. At the same time, few diseases are identical in shapes however vary in color, whereas a few others are identical in color and vary in the shapes. Next to the segmentation of affected and healthy portions, the features relevant to the diseases could be extracted.
To detect the leaf diseases, human vision based models are commonly employed in a traditional way. It requires more time and highly expensive. The accurateness of the human vision model is based on the perception of the persons or experts. Machine learning (ML) and deep learning (DL) approaches allow identification of different kinds of diseases, formulate proper decisions, and choose effective treatment. Only a few works have concentrated on the recognition of rice plant diseases using ML and DL models. Presently, owing to the capability of extracting optimal features, CNN models have been employed comprehensively in ML and pattern recognition researches. Reference [3] reported that the multilayer neural network (NN) offers better learning capability and it denotes the actual data appropriate for classification.
CNN offers dedicated learning solutions and extract related high level features straightaway from the input images. The CNN architectural models are based on the visual cortex of cats in Hubel's and Wiesel's earlier works. Particularly, [4] performs the object detection task by the use of deep CNN. It has tracked by the appearance of numerous enhanced approaches and application areas of CNN. In general, the human recognition of plant diseases is naked eye examination of expert that seems to be timeconsuming and costly in real-time. It is hard to process and also occasionally it generates an error in the identification of the diseases [5]. Owing to the ignorance of appropriate management for rectifying the rice plant leaves diseases, rice productivity is being decreased recently. For resolving this, an appropriate and accurate rice plant detection model is required. Therefore, this paper develops a novel methodology to detect and categorize the diseases present in rice plant images.
This study proposes effective deep learning (DL) based detection and classification model for rice plant diseases using the DenseNet model and multilayer perceptron (MLP), called the DenseNet169-MLP model. In this work, DenseNet model is chosen due to the following advantages: alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. The proposed model comprises set of processes such as preprocessing, fuzzy c-means based segmentation, DenseNet169 based feature extractor, and MLP classifier. The MLP model is integrated with the deep features extracted from the DenseNet169 model for investigating and verifying the efficiency of the deep features of the DenseNet169 model.
The remaining sections of the study are arranged here. Section 2 briefs the works related to the proposed model. Section 3 introduces the rice plant detection and classification model. Section 4 validates the work and Section 5 draws the conclusion.

Literature Survey
This section reviews the plant disease detection models exist in the literature. In [6], a DL based CNN model is applied to detect the three rice plant diseases in an automated way. In the proposed method, the healthy and infected leaf images are differentiated from a total of 1500 rice images. In [7], a new rice plant disease detection and classification model has been presented. Initially, the rice plant images are acquired by a digital camera and then centroid feeding based K-means clustering is applied for segmenting the images. Then, color, shape, and texture features are extracted. Lastly, Support Vector Machines (SVM) is applied for multiclass classification. The presented model has attained 93.3% and 73.3% accuracy on the training and test data correspondingly.
In [8], a method of recognizing and classifying paddy leaves diseases by the use of Deep Neural Network with Jaya Optimization Algorithm (DNN-JOA) is presented. At the preprocessing stage, the background is removed and the RGB image is transformed into the HSV image. Next, the clustering process is performed to segment the affected and healthy portions of the rice plant images. The classification of diseases takes by the use of DNN-JOA. In [9], an ML-based rice plant disease diagnosis model to identify blast disease in South India has been presented. The presented model can identify blast diseases and minimize crop losses effectively. Reference [10] designed a model using DCNN to automatically recognize and categorize biotic and abiotic paddy crop diseases by the use of images captured from the field. The presented model makes use of a pretrained VGG-16 CNN approach to automatically classify the diseased paddy crop images which are gathered at the time of the booting growth stage. In [11], a rice blast detection model using CNN has been presented and it yields better detection rate.
Reference [12] devised a deep machine vision based model to detect, classify and quantify 8 soybean plant stresses, from a total of 25,000 images of stressed and healthy leaves. This method has achieved a detection rate of 94.13% by the presented deep CNN approach. Reference [13] applied 5 distinct CNN approaches to identify plant diseases by the use of leaf images. Reference [14] designed an automated wheat disease identification model utilizing the supervised DL model. It is tested using a wheat image dataset comprising of 50k labeled leaf images. The proposed model includes four diverse CNN models for recognizing the seven wheat disease classes. The results stated that the VGG-16 model has attained a higher detection rate of 97.95%. Reference [15] developed a rice plant disease recognition model by the use of DL approaches. It is tested using 500 leaf and stem images of rice plants. It makes use of LeNet-5 and AlexNet CNN models. This study revealed that the stochastic pooling improvises the generalization capability of the CNN approach and avoids overfitting.
Based on CNN, an effective rice plant identification model has been designed. It is also tested a dataset comprising 500 images of affected and unaffected leaf and stem of rice plants. The proposed method has classified a set of 10 rice diseases and has obtained better accuracy over the traditional ML models. For evaluating the region of interest (RoI), neutrosophic logic technique is employed [16] to determine the presence of infected regions in the image. Different classification models are applied and the random forest (RF) model has attained better detection performance on the applied set of 400 leaf images.
Reference [17] introduced a disease detection model for rice plants. This method determines the diseased area and then calculates the severity level of the disease. Depending upon the severity level, the pesticides are applied for rice plant diseases. In [18], a new rice plant disease detection method has been developed using the Naive Bayes (NB) classification model. This method has detected and categorized three major kinds of rice plant disease and it has less computation time. Reference [19] developed a recent technique for automatic recognition of rice plant diseases. Once the features are extracted, the classification process is carried out using SVM, NB, back propagation neural network (BPNN) and K nearest neighbors (KNN). Reference [20] examined the performance of the diverse transfer learning methods for DNN based plant classification. Reference [21] make use of CNN to detect weeds in soybeans and categorize the weed between grass and broad leaves. A new image dataset is generated comprising 15000 images. The applied CNN model has exhibited better performance on the detection process compared to existing methods. Though several rice plant disease classification models are available in the literature, there is still a need to improve the detection rate in the rice plant disease diagnosis.

The Proposed DenseNet169-MLP Model
The overall workflow involved in the DenseNet169-MLP technique is shown in Fig. 1. The DenseNet169-MLP model undergoes preprocessing to improve the image quality, FCM based segmentation to detect the infected portions, DenseNet169 model for the extraction of the features, and MLP based classification to categorize the images into a different set of class labels. These processes are clearly explained in the subsequent sections.

Preprocessing
Image pre-processing is the procedure employed to eliminate the existence of different noise and other objects present in the image. In the earlier stage, the RGB color image of the rice plant undergoes a channel separation process where the three channels R, G, and B are separated from one another. Next, grayscale conversion process takes place where the RGB rice plant image is transformed into grayscale. Afterward, MF is used to boost the quality of the rice plant image and it is a type of non-linear filtering approach which is commonly applied to remove the noise that exists in the image. The MF technique generally replaces every pixel exists in the input image by the median of the grayscale value of nearby pixels. It is also termed as smoothing spatial filter. Once the filtration task gets done, image segmentation using the FCM technique take place to identify the infected or healthy portions of the rice plant image.

FCM Based Segmentation
Segmentation is considered as an essential process in pattern recognition and image processing based applications. In rice plants, the segmentation process can effectively identify the diseased portions in an automated way. The proposed model make use of FCM based segmention technique and effectively detect the diseased portions from the rice plant leaf images. FCM is commonly employed for soft clustering [22]. It is intended to segment the image X = {x 1 , x 2 …x n }, with n pixels, into c segments (two segments namely diseased and healthy portion), such that each pixel falls into a class with a particular membership degree. The fuzzy set (D F Þ is mathematically represented by: where μ D (x ij ) → [0, 1] defines membership of pixel x j to a certain cluster i in set D. The μ D (x ij ) of every individual pixel. The FCM can be represented by the following objective function as given in Eq. (2): where c denotes the cluster count, n is the pixel count, μ ij is the degree of membership with the limitation such that P C i¼1 m ij ¼ 1, m defines the variable to manage the fuzziness of the clustering procedure and d ij ¼ kx j À # i k is the Euclidean distance among the pixels that exist in the image and cluster prototype. FCM has begun by the initialization of the arbitrary values of membership matrix ranges from 0 to 1, subject to membership constraint. The cluster prototype undergo arbitrary initialization. The minimization of the FCM objective function of FCM takes place by the use of the Lagrange multiplier. The cluster prototype (# i Þ and the memberships (μ ij ) are defined using Eqs. (3)-(4) respectively.

Densenet169 Based Feature Extraction
In this study, the DenseNet169 model is applied to extract the valuable features from the segmented rice plant image. Fig. 2 illustrates the architecture of DenseNet. DenseNet is a network architecture in which the layers are straightaway linked to each other in the dense block in a feed forward way. The layers are defined in a way that the activation map of every earlier layer is treated as individual input whereas its activation map is fed as input to every subsequent layer. In DenseNet [23], the features are concatenated to one another before fed into the layer. Therefore, the lthe layer has l inputs, comprising the feature map of every preceding convolution block. Then, the feature maps are fed to every L − 1 succeeding layer. It establishes LðLþ1Þ 2 links in an L-layer network, as an alternative of just L, as in conventional architectural models. Owing to the dense connectivity pattern, it is referred to as DenseNet. DenseNet comprises a convolution, dense block, transition, and classifier layers after the global average pooling at the input end.

Transfer Learning
Transfer learning is performed on pretrained DenseNet-169 for rice plant disease classification. In contrast to fine tuning process, in this study, feature extraction process is done by eliminating the last fully connected layer (classifier layer) of the DenseNet model and considering the output data of the second final layer as extracted features (termed as DenseNet codes), which are generally high dimension vectors and denotes the features of the input images. Then, the MLP is placed as the final layer of the DenseNet model to perform rice plant disease classification using the extracted features. In the entire process, the DenseNet model is applied as the feature extraction technique and has shown effective outcomes over the existing models [24].
MLP model comprises three components namely input, hidden and output layers. The MLP model contains many diverse hidden layers allowing the network to have computation capabilities of generating the network output. Fig. 4 depicts the architecture of MLP model with an individual hidden layer containing some weights linking among the layers to classify the rice plant image into three classes. The output values would be determined as follows. Initially, the sum of the weights is determined as given in Eq. (5): where x i is the input parameter, w ij is the weight among the input parameter x i and neuron j, and β i is the input parameter's bias term. Next, the output value of the neuron in the hidden layer is created from the received value of weighted summation utilizing a sigmoid activation function as given below.
where f j is the sigmoid function for neuron j and S j is the sum of weight. At last, the outcome of the neuron j can be determined by: where y j is the outcome of neuron j, w ij is the weight among the output parameter y i and neuron j, f j is the activation function for neuron j, and β i is the bias term of the output parameter. The outcome of the MLP represents the appropriate class labels of the applied rice plant images.   Tab. 4 represents the analysis of the comparative results of the DenseNet169-MLP model with recently proposed models [7][8][9] interms of different measures. Fig. 9 shows the sensitivity and specificity analysis of the DenseNet169-MLP model with compared methods. The figure portrayed that the ANN model is found to      figure, it is clear that the ANN approach is said to be an inferior performance that has reached lower precision and F-score of 60.87% and 68.31% respectively. In line with this, the nearer classification function to the ANN method is accomplished by the DAE approach with the precision and F-score of 67.58% and 77.03% correspondingly. Along with that, the KNN scheme has showcased better precision and F-score of 72% and 65% respectively. On the other hand, the DNN technology has attained gradual classifier results with precision and F-score of 74.91% and 81.51% respectively. Moreover, the DNN-JOA technique has ended with considerable results with the precision and F-score of 81.24% and 88.74% respectively. Also, the SIFT-SVM and SIFT-KNN methodologies have referred competing results with the higher precision and F-score values of 86.66%, 86.66% and 90.60%, 90.14% correspondingly. Hence, the   The improved performance is due to the inherent advantages of the DenseNet and MLP models. Therefore, the DenseNet169-MLP model can be employed as a proficient method for rice plant disease diagnosis in a real-time environment to assist farmers and improve crop productivity.

Conclusion
This study has presented an effective DL based DenseNet169-MLP model to classify rice plant diseases. The DenseNet169-MLP model initially had undergone preprocessing to improve image quality. Then, the FCM based segmentation process takes place to detect the infected portions and the DenseNet169 model is applied to extract the features. Finally, MLP is placed at the last layer of the DenseNet169 model to perform final rice plant classification. The experimental validation of the DenseNet169-MLP model takes place on benchmark rice plant dataset. The experimental values verified that the DenseNet169-MLP model has attained maximum sensitivity of 96.40%, specificity of 98.27%, precision of 96.82%, accuracy of 97.68% and F-score of 96.43%. The proposed technique is applied as a proper tool to identify and classify rice plant diseases. In the future, the detection performance can be increased by the use of hyperparameter tuning techniques to fine-tune the DL models.
Funding Statement: The authors received no specific funding for this study.

Conflicts of Interest:
The authors declare that they have no conflicts of interest to report regarding the present study.