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%.
Agriculture acts as a vital part to obtain food security, alleviate poverty, and strengthen the development. Since the global population is anticipated to be 9.7 billion in 2050 and 11.2 billion at the end of this century [
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 [
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 [
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, [
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.
This section reviews the plant disease detection models exist in the literature. In [
In [
Reference [
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 [
Reference [
The overall workflow involved in the DenseNet169-MLP technique is shown in
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.
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 [
where μD(xij) → [0, 1] defines membership of pixel xj to a certain cluster i in set D. The μD(xij) of every individual pixel. The FCM can be represented by the following objective function as given in
where c denotes the cluster count, n is the pixel count, μij is the degree of membership with the limitation such that
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 (
In this study, the DenseNet169 model is applied to extract the valuable features from the segmented rice plant image.
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
In this paper, DenseNet-169 is applied which comprises 4 dense blocks and total 169 layers (165- conv + 3-transition + 1-classification). The layered structure of the DenseNet 169 model is shown in
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 [
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.
In this section, the experimental validation of the DenseNet169-MLP technique has been examined under diverse aspects. The experiments were carried out on a PC i5-8600k processor, GeForce 1050Ti 4GB graphics card, 16GBRAM, and 1TB HDD. The proposed model is simulated using Python 3.6.5 tool. The parameter setting is given as follows: mini batch size: 200, dropout: 0.5, number of hidden layers:3, number of hidden units: 1024, and learning rate: 0.01. The measures used to analyze the results of the DenseNet169-MLP model are Sensitivity, Specificity, Precision, Accuracy, F-score, Negative Predicted Value (NPV), False Positive Rate (FPR), False Negative Rate (FNR) and False Discovery Rate (FDR). For validation, a benchmark rice plant image dataset is used [
For better understanding, the confusion matrix in the figure is transformed into a tabular form as given in
Input labels | Class labels | Total images | ||
---|---|---|---|---|
Bacterial leaf blight | Brown spot | Leaf smut | ||
Bacterial leaf blight | 40 | 0 | 0 | 40 |
Brown spot | 0 | 33 | 4 | 37 |
Leaf smut | 0 | 0 | 38 | 38 |
Total images | 40 | 33 | 42 | 115 |
Different levels | Bacterial leaf blight | Brown spot | Leaf smut |
---|---|---|---|
TP | 40 | 33 | 38 |
TN | 75 | 78 | 73 |
FP | 0 | 0 | 4 |
FN | 0 | 4 | 0 |
Measures | Bacterial leaf blight | Brown spot | Leaf smut | Average |
---|---|---|---|---|
Sensitivity/TPR | 100 | 89.19 | 100 | 96.40 |
Specificity/TNR | 100 | 100 | 94.80 | 98.27 |
Precision/PPV | 100 | 100 | 90.47 | 96.82 |
Accuracy | 100 | 96.52 | 96.52 | 97.68 |
F-score | 100 | 94.29 | 95.00 | 96.43 |
NPV | 100 | 95.12 | 100 | 98.37 |
FPR | 0 | 0 | 5.19 | 1.73 |
FNR | 0 | 10.81 | 0 | 3.60 |
FDR | 0 | 0 | 9.52 | 3.17 |
Finally, the DenseNet169-MLP model has classified the leaf smut images with the sensitivity of 100%, specificity of 94.80%, precision of 90.47%, accuracy of 96.52%, F-score of 95%, NPV of 100%, FPR of 5.19%, FNR of 0 and FDR of 9.52%. These values demonstrated the proficient characteristics of the DenseNet169-MLP model on the classification of rice plant diseases.
Methods | Sensitivity | Specificity | Precision | Accuracy | F-score |
---|---|---|---|---|---|
DenseNet169-MLP | 96.40 | 98.27 | 96.82 | 97.68 | 96.43 |
VGG-16 CNN (2020) | - | - | - | 92.89 | - |
DNN_JOA (2019) | 83.70 | 94.04 | 81.24 | 94.25 | 88.74 |
DNN (2019) | 73.46 | 89.42 | 74.91 | 90.00 | 81.51 |
DAE (2019) | 68.02 | 87.18 | 67.58 | 86.04 | 77.03 |
ANN (2019) | 63.30 | 81.58 | 60.87 | 80.01 | 68.31 |
CNN (2019) | 94.00 | 94.00 | 94.00 | 94.00 | 94.00 |
KNN (2019) | 65.00 | 78.00 | 72.00 | 70.00 | 65.00 |
BPNN (2018) | - | - | - | 95.83 | - |
SVM (2017) | - | - | - | 93.33 | - |
SIFT-SVM (2016) | 86.66 | - | 86.66 | 91.10 | 86.66 |
SIFT-KNN (2016) | 90.00 | - | 90.60 | 93.33 | 90.14 |
PCA-SVM (2015) | - | - | - | 96.55 |
Also, the DNN-JOA model has resulted in somewhat acceptable results with the sensitivity and specificity of 83.70% and 94.04% respectively. Moreover, the SIFT-SVM and SIFT-KNN models have exhibited competitive results with the maximum sensitivity values of 86.66% and 90% respectively. However, the CNN model has a slightly higher result with sensitivity and specificity values of 94% and 94%. At last, the DenseNet169-MLP model has demonstrated superior performance by attaining maximum sensitivity and specificity values of 96.40% and 98.27% respectively.
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.