Guava is one of the most important fruits in Pakistan, and is gradually boosting the economy of Pakistan. Guava production can be interrupted due to different diseases, such as anthracnose, algal spot, fruit fly, styler end rot and canker. These diseases are usually detected and identified by visual observation, thus automatic detection is required to assist formers. In this research, a new technique was created to detect guava plant diseases using image processing techniques and computer vision. An automated system is developed to support farmers to identify major diseases in guava. We collected healthy and unhealthy images of different guava diseases from the field. Then image labeling was done with the help of an expert to differentiate between healthy and unhealthy fruit. The local binary pattern (LBP) was used for the extraction of features, and principal component analysis (PCA) was used for dimensionality reduction. Disease classification was carried out using multiple classifiers, including cubic support vector machine, Fine K-nearest neighbor (F-KNN), Bagged Tree and RUSBoosted Tree algorithms and achieved 100% accuracy for the diagnosis of fruit flies disease using Bagged Tree. However, the findings indicated that cubic support vector machines (C-SVM) was the best classifier for all guava disease mentioned in the dataset.
Agriculture is a major source of food, and provides a significant contribution to the economy [
Guava plants are prone to many diseases. In Pakistan, for example, anthracnose is a major problem. In order to protect the plants from disease, it is important to be able to identify the diseases rapidly and accurately, so the use of an automated detection system has many advantages. Our research focused upon the detection and classification of diseases in guava. Images of agricultural value suffer from several issues, such as the automation of plant disease identification, the identification of multiple diseases, variability in the resolution of the images, partial occlusion of plants due to surrounding vegetation, and variability in illumination and lighting conditions. The challenge is to decrease the numbers of false positives and increase the accuracy of detection and classification of infected regions.
The main aim of this work was to develop an automatic technique for identifying and classifying regions of disease in guava plants. The following steps were performed to detect and classify guava diseases using image processing techniques. Images of four types of guava diseases were collected from the field, and labeled by experts to differentiate between healthy and unhealthy fruits. We used YCbCr and red/green/blue (RGB) color representations in the preprocessing phase to improve the contrast of the original image, facilitating the detection of infected regions in the plant image. In the feature extraction phase, we used the LBP method, which is useful for the improvement of the detection process. The color features were extracted from the segmented regions. Feature level image fusion was used to sharpen image resolution and to improve the classification of the image. Finally, we utilized multiple classifiers including C-SVM F-KNN, RUSBoosted, and bag tree algorithms to classify four diseases regarding guava.
Research into the automated detection of plant diseases has been of interest to researchers for many years. Gavhale and Gawande built a model for the identification of different diseases in plants using images of plant leaves. Their approach involved five stages. In step one, they used a camera to capture initial image sets, and preprocessed these images to enhance the images and color space. For segmentation, the infected regions were identified using edge, region, and threshold-based segmentation techniques, and then texture, color features, and shape were calculated. Finally, a neural network classifier was used to create texture feature taxonomy [
Gavhale et al. developed a framework for the identification of disease affected parts of citrus leaves. They recognized the disease using image preprocessing techniques including image enhancement, RGB color vector transformation, and
Khan et al. [
A comprehensive description of the major diseases of guava is given in [
Anthracnose is a fungal disease. It produces dipped, dark-colored cuts on mature fruit, which may become covered in pink spores. The cuts combine to form large necrotic patches on the surface of the fruit. Fungicides are used to maintain the disease. Fruit affected by anthracnose is shown in
Algal spot is also fungal. It causes spots on the leaves and fruits, which reduce the photosynthetic capacity of the plant. This disease does not produce major economic losses. Fruit affected due to anthracnose is shown in
This disease is caused by infection with the ascomycete fungus
Fruit flies carry a bacterial disease. The symptoms of the disease are depressions in the fruit with dark-colored lesions and soft areas caused by larvae feeding on the fruit. The growth of secondary rots frequently causes fruit to drop from the tree. Fruit affected by fruit fly is shown in
Our proposed technique involves the following five stages: image acquisition, image labeling, feature extraction and fusion, feature reduction, and finally classification.
Images of guava plants were acquired from the field. Infected leaves and fruit were picked from the plant and placed on a white background. While capturing the image, we adjusted the camera to capture the image with a white background for the fruit and leaf only. Our database comprises 400 images covering four kinds of guava diseases: anthracnose, algal spot, styler end rot, and fruit flies, and images of healthy fruit. Each image has dimensions 520 × 530 at 300 dpi.
We labeled the images with the help of an expert. Image labeling was done on all images, whether the plants were healthy or not, based on the symptoms produced by each disease.
Feature extraction involves the transformation of raw data into a set of features. Features play an important role in the field of image processing [
In our methodology, we applied two methods for feature extraction. First, the color features of the image were extracted, and then LBP method was applied to the dataset to obtain matrices of images.
The color features are very useful and more important because each pixel consists of a different color [
The mean was calculated using the following formula:
where
Standard deviation was calculated using the following formula:
where
Skewness was measured by the following formula:
where
Entropy was measured using:
where H(
LBP is a graphic descriptor used in the field of pattern recognition [
where
The resulting features may be very high dimensional in the case when features are fused, and feature reduction algorithms are required to reduce the dimensionality. The traditional approaches usually employed for feature reduction include PCA [
where
The selected features were then fed to the classifiers. Five types of classifier that is KNN, M-SVM, RUSBoosted Tree, F-KNN and Bagged Tree classifier [
Classification produces four values, true positive (TP), true negative (TN), false positive (FP) and false negative (FN) in the form of a confusion matrix. Following classification, different performance measures can be calculated, including accuracy, True Positive Rate (TPR), FN Tate (miss rate), Positive Predictive Value (PPV) and area under the ROC curve (AUC).
Accuracy is the number of TP and TN results out of the total numbers of instances evaluated [
In this section, we present an analysis of the experimental result and review the performance of the approach we developed. In general, the proposed technique focuses on the following five steps i) image collection, ii) labeling, iii) feature extraction, iv) feature reduction and v) classification, as depicted in
Classifier | TPR | PPV | FNR | AUC | Accuracy (%) |
---|---|---|---|---|---|
BT | 94 | 97 | 4.3 | 0.97 | |
F-KNN | 92 | 96 | 5.4 | 0.92 | 94.6 |
C-SVM | 92 | 96 | 5.4 | 0.97 | 94.6 |
W-KNN | 89 | 95 | 7.5 | 0.95 | 92.5 |
RBT | 79 | 85 | 16.1 | 0.98 | 83.9 |
Fruit type | No. of images | Unhealthy (%) | Healthy (%) |
---|---|---|---|
Unhealthy | 40 | 100 | 0 |
Healthy | 40 | 6 | 94 |
Classifier | TPR | PPV | FNR | AUC | Accuracy (%) |
---|---|---|---|---|---|
C-SVM | 97.5 | 98.5 | 2 | 0.99 | |
F-KNN | 94.5 | 97 | 4.1 | 0.95 | 95.9 |
BT | 94.5 | 97 | 4.1 | 0.97 | 95.9 |
W-KNN | 92 | 95.5 | 6.1 | 0.97 | 93.9 |
RBT | 81.5 | 46 | 15.3 | 0.95 | 84.7 |
Fruit type | No. of images | Unhealthy (%) | Healthy (%) |
---|---|---|---|
Unhealthy | 40 | 95 | 5 |
Healthy | 40 | 0 | 100 |
Classifier | TPR | PPV | FNR | AUC | Accuracy (%) |
---|---|---|---|---|---|
C-SVM | 94.5 | 99 | 1.4 | 1.00 | |
F-KNN | 98.5 | 91 | 2.9 | 0.99 | 97.1 |
BT | 93 | 59 | 4.3 | 0.93 | 95.7 |
W-KNN | 87.5 | 87.5 | 5.7 | 0.91 | 94 |
RBT | 61 | 95 | 10.0 | 0.94 | 90 |
Fruit type | No. of images | Unhealthy (%) | Healthy (%) |
---|---|---|---|
Unhealthy | 40 | 89 | 11 |
Healthy | 40 | 0 | 100 |
Classifier | TPR | PPV | FNR | AUC | Accuracy (%) |
---|---|---|---|---|---|
BT | 100 | 100 | 0 | 1.00 | |
C-SVM | 98.5 | 99 | 1.1 | 1.00 | 98.9 |
F-KNN | 98 | 99 | 1.1 | 0.98 | 98.9 |
W-KNN | 92.5 | 96 | 5.3 | 0.98 | 94.7 |
RBT | 74.5 | 45 | 19.1 | 0.97 | 80.9 |
Fruit Type | No. of images | Unhealthy (%) | Healthy (%) |
---|---|---|---|
Unhealthy | 30 | 100 | 0 |
Healthy | 30 | 0 | 100 |
Guava is a product of the tropical areas of the world, and plays an important role in the economy of Asia, especially in Pakistan. The detection of guava plant diseases is an important issue in agriculture. Many classifiers and techniques have been used to detect guava diseases. In this work, we used image processing techniques to detect diseases in guava. An automatic framework was developed to support farmers in the detection and classification of major diseases in guava. We selected four types of guava plant diseases: anthracnose, algal spot, fruit flies, and styler end rot. These images were collected from from the field. The LBP was used for the extraction of features, and principal component analysis was used for dimensionality reduction. The techniques achieved accuracies of 95.7%, 95.9%, 94%, 95.7%, and 100% were achieved for anthracnose, algal spot, styler end rot, and fruit flies respectively using a Bagged Tree classifier. The proposed methodology provides efficient results. The assessment of the classifiers for large datasets, using the present methodology can be attempted in future.
The authors extend their appreciation to the Deanship of Scientific Research at King Saud University for funding this work through research Group No. RG-1441-379.