Learning Rate and Epoch Analysis for Medicinal Plant Identification Using GLCM and BPNN

Abstract

Accurate identification of medicinal plants is essential for pharmacology and biodiversity conservation. However, traditional methods rely heavily on subjective visual inspection, which is prone to misclassification due to subtle differences in leaf textures. A primary challenge that remains unaddressed is the understanding of hyperparameter sensitivity within limited datasets, particularly when the subjects exhibit extremely high visual similarity. This study proposes an automated identification approach using Gray-Level Co-occurrence Matrix (GLCM) and Backpropagation Neural Network (BPNN) to classify three Indonesian medicinal species: white ginger, mango ginger, and yellow turmeric. The distinctive focus of this research lies in its attempt to differentiate these specific plants, which possess leaf texture characteristics so similar that they are often indistinguishable to the human eye. This approach involves a systematic analysis of learning rate and epoch parameters to optimize convergence for these nearly identical texture features. A dataset of 63 images was transformed into five GLCM statistical features to serve as the primary inputs for the BPNN. Experimental results demonstrate that classification performance is highly sensitive to parameter tuning. The system achieved its peak accuracy of 65.03% using a learning rate of 0.1 and 100 epochs. The findings reveal that smaller learning rates and limited training iterations facilitate more stable convergence when processing data with high feature similarity. While the accuracy indicates potential for further development, this study provides a significant contribution to creating objective identification methods for visually similar plants and offers empirical insights into optimal parameter selection for texture-based neural network architectures.