Plant Disease Detection and Classification Using Artificial Neural Network
No Thumbnail Available
Date
2019-07
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Continuous advancement in image processing and machine learning techniques have made it
possible for computers to see and learn. What is seen by the eyes of human beings could be divided
into pixels and given to a computer so that the computer will be able to see and learn based on the
provided values. Based on the input values fed in computers could learn to identify various things
based on the things they have been learnt from them. There are many possible areas in which
computers can be applied to see and learn in order to make the life of human beings much easier.
In this thesis an approach has been proposed which is capable of automatically detecting and
classifying plant disease from an image based on artificial neural network. Now a days, plants have
become much more important than they used to some years ago where they have been only used
to feed mankind as well as animals. Plant diseases are currently detected and classified using
methods that requires a lot of manual work with experts, agricultural extension worker and farmers
which is both time consuming and error prone. To automate the process of plant disease detection
and classification different researchers have studied many techniques using both machine learning
and image processing. However, these proposed techniques still have limitation.
The steps followed in this research for detecting and classifying the plant disease are: dataset
collection, image pre-processing, masking, and removing the green part, feature extraction and
selection, classification, and disease management techniques. For comparing and demonstrating
the conventional machine learning techniques and proposed approach respectively two different
types of plant have been selected namely, maize, and potato from the plantvillage.org website.
Since the conventional machine learning techniques do not have the potential to extract and select
features from a given raw data, texture features using Haralick’s from color co-occurrence matrix
have been extracted and selected using subset feature selection technique.
The proposed approach and the selected conventional machine learning techniques were evaluated
using confusion matrix, classification performance report, and t-test to asses which has the higher
classification potential. The proposed approach achieved an average accuracy of 97.6%, average
precision of 97.0%, 97.0% of average recall ,and average F1 value of 97.0% over a test dataset of
previously unseen 1201 images. From the analysis of the experimental results the proposed
approach gives best result than the conventional machine learning classifier. This due to the fact
that convolutional neural network extract high level features from the input raw data, making it
more efficient and accurate, and avoid errors due to a subjective manual feature extraction thereby
showing the feasibility of its usage in real time applications for the classification of healthy and
non-healthy plants.
Description
Keywords
plant disease, image processing, masking and removing feature extraction and selection, machine learning methods, disease management techniques