Identification of Philippine Maize Variety Using Convolutional Neural Network with Kernel Morphological Phenes Characterization

Abstract

Maize variety significantly increased due to hybridization and the characteristics of individual seed overlap causing the non-uniformity problem in agricultural production. Non-destructive optical techniques already exist but some have high operation requirements. This study developed a nondestructive RGB imaging model for classifying Philippine maize varieties, namely IPB VAR 6, NSIC CN 282, and PSB CN 97–92, using a 17-layer convolutional neural network (CNN) with 7.040858x106 extracted parameters. CIELab thresholding was performed to extract the morphological phenes (roundness, compactness, solidity, shape factors) that were also used as input predictors in feature-based predictive machine learning modeling. The developed CNN model outperformed the optimized decision tree, Naïve Bayes, linear discriminant analysis, k-nearest neighbors, and MobileNetV2 models based on the accuracy (96.667%), sensitivity (96.667%), and specificity (96.970%) in discriminating 300 kernels provided by the Philippine Bureau of Plant Industry. Based on the morphological phenotyping, the IPB VAR 6 has the most prominent shape factors 1 and 2 which are based on major and minor axis lengths of the kernel. NSIC CN 282 is the roundest, most compact, and has significant shape factors 3 and 4, while PSB CN 97–92 has the highest average solidity value. This developed approach has a great advantage for in situ classification and phenotyping without requiring performing laboratory procedures and shelling out huge expenses for the technology.