Herbs recognition system based on physiochemical properties using weighted histogram and multiple discriminant analysis methods

Currently, herbs recognition system has become a promising method to identify herbs species. Misuse of herbal medicine can cause serious health problems due to toxicological effects of phytochemical. As a result, a system that able to distinguish the types of herbs is needed. Most herbs recognition systems available in the market are dependent on experts. In this research, the concern is to identify the herbs compounds within the same group species where the physical appearance and aroma are similar. The work mainly focuses on herbs recognition systems that intended for researchers and medical practitioners use without the need for experts. The study is mainly based on chemical and physical properties as well as the combination of both. Hence, three feature extraction methods for herbs recognition system based on chemical and physical properties of herbs are presented. Electronic Nose (E-Nose) devices have been used extensively to differentiate and characterize the herb species based on their unique odor. Electrical signal generated from the gas sensor array is one of the physical properties studied. Then, a Gas Chromatography-Mass Spectrometry (GCMS) device is utilized to extract the chemical compound of herbs. The first feature extraction technique in this research is Principal Component Analysis (PCA) and selected feature based on electrical signal, however it is an unsupervised learning. Thus, Multiple Discriminant Analysis (MDA) is proposed as the second feature extraction technique. MDA is one of the supervised learning techniques to replace PCA. The third feature extraction technique is proposed to develop an automated GCMS system that differentiates the herbs species from the major volatile compounds. The Weighted Histogram Analysis Method (WHAM) is proposed to make use of both major and minor volatile compounds in GCMS herbs recognition system. Fusion techniques have been extensively studied on multisensory environments. Comparison between system with and without feature fusion techniques is presented. In this research, 19 herbs species from 5 family groups are studied. The robustness test of the three proposed herbs recognition systems are performed via four classification models: Support Vector Machine (SVM), KNearest Neighbors (KNN), Multinomial Logistic Regression (MLR), and Gaussian Radial Basis Function (RBF) Kernel. In GCMS, the overall system performance with WHAM improves the classification accuracy by average of 0.6%-38.34% using SVM and KNN. In E-Nose herbs recognition system, an average between 0%-21.73% improvement in system performance with MDA. The performance of classification accuracy using KNN shows a better result within the family group from 92.15% to 100% compared to the other methods. In addition, KNN method shows classification accuracy improvement in average of 99.58%-100% within the same family group. As a conclusion, the system with KNN method is capable to classify herbs species more accurately as compared to the other three classification method. The innovation of this research could benefit especially the researchers, to identify the plant species without relying on the expertise of botanists and forest rangers for the learning and training process.

Text NUR FADZILAH BINTI MOHAMAD - IR.pdf Download (1MB)

Actions (login required)

View Item