HYBRID DEEP LEARNING WITH MULTIMODAL DATA AUGMENTATION FOR RIVERBANK TREE RING CLASSIFICATION

Abstract

The purpose of this research was to develop a model for counting or classifying annual tree rings of lamphu trees along riverbanks. The study was based on image processing combined with deep learning and machine learning to develop the model. The image data augmentations were processed through image preprocessing with the region of interest, denoising, color model conversion, blurring, thresholding, and dilation. This process transforms one image’s data into seventy images as the dataset. Deep learning models developed with Visual Geometry Group-16-based convolutional neural networks applied different support vector machine classifiers: L2 norm, categorical hinge loss, and the Weston-Watkins. The results showed that the model developed with the categorical hinge loss-support vector machine classifier gave the highest model efficiency with an accuracy of 94.07% and a categorical cross-entropy loss of 5.48%. The models developed with the Weston-Watkins and L2 norm were followed by an accuracy of 93.70% and 93.26%, respectively. The model developed with the Softmax function as a classifier that gave the lowest order of accuracy of 92.96%. Keywords: Convolutional Neural Network, Data Augmentation, Deep Learning, Support Vector Machine, Tree Ring DOI： https://doi.org/10.35741/issn.0258-2724.58.3.27