Plant Pathology Detection Using Image Processing

S. Syed Athaullah, S. Hari Prasath, S.M.Giri Rajkumar,S.K. Syed Fizal,

I. Shanjith Kumar,G. Ragavantiran .

Department of Instrumentation and Control Engineering , Saranathan College of

Engineering, Trichy.

Abstract.

Introduction

About 70% of the population in India is dependent on agriculture. For the purpose of preventing
yield losses, it is crucial to identify plant illnesses. Manually observing the plant illnesses is very
difficult. It requires a significant amount of work, knowledge of plant diseases, and a significant
amount of time. In order to identify plant diseases, image processing algorithms can be used. In
this project, we have outlined a method for identifying plant diseases using images of the leaves.
Image processing is a subset of signal processing that can extract valuable information or image
properties. A component of artificial intelligence called machine learning performs tasks
automatically or provides guidance on how to carry them out. Understanding the training data
and incorporating it into models that should be helpful to people is the primary goal of machine
learning. Consequently, it can help in making wise choices and predicting the right output using
the vast training data. For classification, factors like leaf hue, leaf damage severity, leaf area, and
leaf texture are taken into account. In order to identify various plant leaf diseases with the
greatest precision, we have examined various image parameters or features. In the past,
specialists used chemical processes or visual inspection of the leaves to identify plant diseases.
This requires a sizable team of experts and ongoing plant observation, both of which are
expensive when done with big farms. The suggested method works well in these circumstances
for keeping an eye on vast fields of crops. It is simpler and less expensive to automatically
identify diseases by simply observing the symptoms on plant leaves. Since the suggested
approach employs image processing algorithms, it is computationally less expensive and takes
less time to predict based approaches.

Existing Methods

Using back propagation neural network (BPNN) technology and digital image processing
methods, S. Khirade et al. addressed the issue of plant disease detection [1]. Authors have
developed various methods for spotting plant diseases from pictures of the leaves. To segment
the infected portion of the leaf, they used Otsu's thresholding, boundary detection, and spot
recognition algorithms. Then, for the purpose of classifying plant diseases, they extracted
characteristics such as color, texture, morphology, edges, etc. For categorization, or to find the
plant disease, BPNN is used.
In their study, Shiroop Madiwalar and Medha Wyawahare examined various image processing
strategies for plant disease identification [2]. The ability to identify plant disease using colour
and texture features was examined by authors. On the collection of 110 RGB images, they tested
their algorithms. The GLCM features, the mean and standard deviation of the picture convolved
with the Gabor filter, and the mean and standard deviation of the RGB and YCbCr channels were
the features extracted for classification. For categorization, a support vector machine classifier
was employed. The authors came to the conclusion that GCLM characteristics are useful for
spotting healthy leaves. While colour features and Gabor filter features are thought to be the best
for spotting leaf spots and anthracnose-affected foliage, respectively. Using all the extracted
features, they were able to obtain the highest accuracy of 83.34%.
The use of hyperspectral imaging in the task of identifying plant diseases was proven by Peyman
Moghadam et al. [3]. This study made use of the visual, near-infrared (VNIR), and short-wave
infrared (SWIR) spectrums. For the segmentation of leaves, authors used the k-means clustering
method in the spectral domain. For the purpose of removing the grid from hyperspectral images,
they have suggested a novel grid removal algorithm. The accuracy of the authors' vegetation
indices in the VNIR spectral range was 83%, and their complete spectrum accuracy was 93%.
Despite the fact that the suggested technique produced higher accuracy, it is too expensive
because it calls for a hyperspectral camera with 324 spectral bands.
The Bacterial Blight Detection System for Pomegranate Plant was created by Sharath D. M. et al.
using features like color, mean, homogeneity, SD, variance, correlation, entropy, edges, etc.
Grab cut segmentation has been used by the authors to separate the area of interest in the image
[4]. The edges of the pictures were extracted using Canny edge detector. A system that can
accurately forecast the level of fruit infection has been created by the authors.
Convolutional neural networks were used by Garima Shrestha et al. to identify the plant illness
[5]. With 88.80% accuracy, authors were able to effectively classify 12 plant diseases. The
experimentation used a dataset of 3000 high definition RGB images. Three sets of convolutional
and pooling layers make up the network. This increases the network's processing cost.
Additionally, the model's F1 score is 0.12, which is extremely low due to the high amount of
false negative predictions.

Proposed System

Dataset

For this study, Sharada P. Mohanty et al.'s PlantVillage public dataset for plant leaf disease
detection was used [6]. There are 38 groups in the dataset's 87,000 RGB images of healthy and
unhealthy plant leaves, but we've only chosen 25 of them to test our algorithm on. Table 1
displays these groups.

Table 1. Dataset Specifications

Plant Disesase Name No of Images

Apple Healthy 2008

Diseased Scab 2016
Diseased : Black rot 1987
Diseased: Cedar apple 1760
rust

Corn Healthy 1859

Diseased : Cercospora 1642
 leaf spot
Diseased : Common 1907
rust 1908
Diseased : Northern
Leaf Blight

Grapes Healthy 1692

Diseased :Black rot 1888
Diseased : Esca (Black 1920
Measles)
Diseased :Leaf 1722
Blight(Isariopsis)

Some dataset samples are shown in table 1.

A. Data Preprocessing and Feature Extraction

A crucial job in any computer vision-based system is data preprocessing. The preprocessing
procedures for each picture are shown in Fig. 2. Before extracting the features, some background
noise should be eliminated to obtain exact results. In order to smooth the picture, a Gaussian
filter is used after the RGB image has been converted to greyscale. The Otsu thresholding
algorithm is then used to binaries the picture. The tiny gaps in the foreground are then filled
using morphological transform on the binarized image. Following the identification of the
foreground, the RGB image of the segmented leaf is obtained by performing a bitwise AND
operation on the binarized image and the original colour image. Shape, texture, and colour
characteristics are now extracted from the image following segmentation. The area and perimeter
of a leaf are determined using contours. The line that connects all of the points along the edges of
things with the same colour or intensity is known as a contour. Each RGB channel's mean and
standard deviation are also calculated. The picture is first converted to HSV colour space, and
the ratio of pixels with hue (H) channel pixel intensities between 30 and 70 and the total number
of pixels in one channel is then calculated to determine the amount of green colour in the image.
Calculating the non-green portion of a picture requires subtracting the green component from 1.
We have extracted texture features from the picture's grey level co-occurrence matrix (GLCM)
after extracting colour features from the image [7].

Original Image

Grayscale
Conversion

Smoothening
Using Gaussian
Filter

Otsu’s
Thresholding

Shape Features Morphological

Transform

Bitwise AND
Colour Features operation with HSV Color
original frame Space
Conversion

Gray level Co-

occurrence matrix

Texture Green part

Features In the leaf

GLCM is the spacial relationship of pixels in the image. Extracting texture features from GCLM
is one of the tradition method in computer vision. We have extracted following features from
GCLM:
• Contrast
• Dissimilarity
• Homogeneity
• Energy
• Correlation

After extracting all the features from all the images in the dataset, feature selection task is
performed.

B. Algorithm Classification

For categorization or detection tasks, the random forest classifier has been used. It is a
component of learning in which several base estimators are used to forecast the output [8].
Typically, decision trees are used to attain higher accuracy. However, they are vulnerable to
overfitting issues. Therefore, the random forest classifier, which is a mix of various decision
trees, is used to address this problem. Therefore, a random forest classifier—a mix of various
decision trees—is used to address this problem. Each tree is trained using various subsets of the
entire dataset, which can lessen overfitting and increase the classifier's precision. The dataset has
been divided into a train set (80%) for model fitting and a test set (20%) for model evaluation.
Utilizing K-fold cross validation, the accuracy value is determined. With no bias, this technique
can determine accuracy across the entire dataset. Following data fitting, the f1 score, precision,
recall, and accuracy have been determined from test data to evaluate the model's effectiveness.
For the analysis of erroneous positives and false negatives, the ROC curve and confusion matrix
were plotted.
Result Analysis

The performance vectors for each model created for each plant are displayed in Table 2. As we
can see, the accuracy values are very similar to the f1 scores. This is due to a balanced amount of
incorrect forecasts, both positive and negative. The ideal scenario for any machine learning
programme is this. The precision was 93% on average.

Plant Accuracy F1 Score

Apple 0.91 0.91

Corn 0.94 0.94
Grapes 0.95 0.95
Conclusion

We have successfully developed a plant pathology detection with average accuracy of 93% and
0.93 F1 score. Also the proposed system is computationally efficient because of the use of
statistical image processing .