HYBRID FEATURE EXTRACTION FOR DETECTION OF PLANT DISEASES

• Main Author: Naufal, Muhammad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51165
• Institution: Institut Teknologi Bandung

With current technological advances, various ways have been developed to detect plant diseases. Diagnosing plant diseases manually requires a lot of money and time. Therefore, various approaches have been developed, one of which is to use machine learning methods. This method utilizes feature extraction to detect diseases from plant images. There are several feature extraction that will be compared, namely HOG (Histogram of Oriented Gradients), ORB (Oriented FAST and Rotated BRIEF), and SURF (Speeded-Up Robust Features). In addition, an experiment will be conducted to combine and compare the feature extraction (hybrid feature extraction). The main objective of this experiment is to find the best hybrid function, based on the accuracy metric, F1-Score, and Area Under ROC. The dataset used comes from PlantVillage. The data type in this dataset is an image in JPEG format. This dataset consists of three different classes, namely potato early blight, potato late blight, and healthy. The classifiers to be used are SVM (Support Vector Machine) and Decision Tree. The test is evaluated with the metrics for accuracy, F1-Score, and Area Under ROC (AUROC). Based on the tests that have been done, it was found that the hybrid feature extraction method of ORB and SURF (ORB + SURF) and the SVM classifier produced predictions with an accuracy of 89.87% and an F1-Score of 89.77%. The hybrid feature extraction ORB + SURF and the SVM classifier also have the best Area Under ROC value, which is 0.93.