| Title |
Enhanced CNN-based Plant Growing-stage Classification using Additional Information Carried in an Additional Channel |
| Authors |
Yeh-Huann Goh(Yeh-Huann Goh) ; Zhao-Jun Ho(Zhao-Jun Ho) ; Chee-Ho Ng(Chee-Ho Ng) ; Yann-Ling Goh(Yann-Ling Goh) |
| DOI |
https://doi.org/10.5573/IEIESPC.2019.8.3.171 |
| Keywords |
Plant phenotype ; Convolutional neural network ; Additional image channel ; Plant growing stage |
| Abstract |
Studying the observable characteristics of mutants and the growing stages of the same genotype plant interacting with various environmental conditions is important in order to understand the performance of a particular trait in different growth environments. A plant's growing stage affects the growth rate of leaves, the photosynthetic rate, water absorption capacity, and other characteristics. By automating the plant mutant classification process and the growing-stage classification process, botanists and agriculture scientists can perform large-scale experiments to cultivate plants with useful traits to combat extreme environmental conditions. This research aims to construct an enhanced optimum convolutional neural network (CNN) for image-based plant growing-stage classification, and a description of the algorithms to construct the optimum CNN for image-based plant mutant classification is included as well. This research was carried out using the Ara2013-Canon dataset annotated by the International Plant Phenotyping Network (IPPN) for classification processes. Optimum parameters found in this paper are for 1) the number of convolutional layers, 2) the number of neurons in a fully connected (FC) layer, and 3) the number of FC layers in a CNN for plant growing-stage classification. The possibility to enhance the successful classification rate is explored by introducing an additional channel that carries additional useful information, such as 1) mutant type, 2) number of leaves, 3) total size for all leaves, 4) mean leaf size, and 5) standard deviation of the leaf size in the form of square matrices. Experimental results show that under optimum conditions for growing-stage classification, a CNN classification system utilizing plant images and all additional useful information provides the best recognition rate at 81.97%. |