Deep learning-based approach in plant species identification / Tan Jing Wei

Tan, Jing Wei (2018) Deep learning-based approach in plant species identification / Tan Jing Wei. Masters thesis, University of Malaya.

[img] PDF (The Candidate's Agreement)
Restricted to Repository staff only

Download (195Kb)
    PDF (Thesis M.A)
    Download (2350Kb) | Preview


      Plant species identification and classification is one of the main tasks for botanists as well as a matter of interest for public. An automated plant species identification system could help the botanists and the layman to identify plant species in a more structured and speedy manner. Conventional machine learning techniques are widely used in the development of automated identification system in various fields including in biology and biodiversity. Deep learning is an emerging area in the machine learning approach. It has been considered as one of the powerful approaches for feature extraction as compared to the conventional approaches due to its superiority in providing deeper information of an image rather than the surface information. In this research, a total of 1290 leaf samples were collected in the University of Malaya (UM), Malaysia from 43 species of tropical trees with 30 samples for each species. The leaf images were pre-processed based on the feature extraction approaches which included the removal of background noises, segmentation of region of interest (ROI) and conversion of RGB images into grey-scaled images. The features were then extracted by using one of the deep learning approaches which is Convolutional Neural Network (CNN). Based on the literature review, this is one of the first few studies, which has applied CNN in tropical tree species identification, by using both leaf morphometric and venation pattern approaches. Three CNN-based models were used for feature extraction which are pre-trained AlexNet, fine-tuned pre-trained AlexNet and a newly proposed CNN model – D-Leaf model. A conventional morphometric method was employed for benchmarking purposes, which computed the morphological measurements based on the Sobel segmented veins. These features were classified by using four machine learning techniques, namely, Support Vector Machine (SVM), Artificial Neural Network (ANN), k-Nearest Neighbour (k-NN), Naïve Bayes (NB) and Convolutional Neural Network (CNN). The fine-tuned AlexNet model performed slightly better (testing accuracy = 95.54%) than the D-Leaf (testing accuracy = 94.88%) models and AlexNet (testing accuracy = 93.26%). However, the execution time of D-Leaf model was 7 times faster than AlexNet and fine-tuned AlexNet models, respectively. The CNN models obtained a much higher performance than the vein morphometric measurement model which obtained only 66.28% in testing accuracy. In addition, ANN classifiers have achieved much better performance than SVM, k-NN, NB and CNN. In this research, D-Leaf can be a more efficient and effective automated tool for plant species identification with a high accuracy and shorter execution time than AlexNet and the fine-tuned AlexNet models as the CNN models performed better than the conventional morphometric measurements model. The conventional morphometric measurements method was less desirable in extracting features as compared to the CNN approach. The CNN extracted features are found to be fitted well with the ANN classifier as compared to other classifiers.

      Item Type: Thesis (Masters)
      Additional Information: Thesis (M.A.) – Faculty of Science, University of Malaya, 2018.
      Uncontrolled Keywords: Tropical plants; Deep learning; Convolutional Neural Network; Artificial Neural Network
      Subjects: Q Science > Q Science (General)
      Divisions: Faculty of Science
      Depositing User: Mr Mohd Safri Tahir
      Date Deposited: 05 Jan 2021 02:58
      Last Modified: 14 Jan 2021 08:50

      Actions (For repository staff only : Login required)

      View Item