Innovative agricultural technologies increasingly utilize artificial intelligence (AI) and machine learning to enhance productivity and precision. Among these advancements, Convolutional Neural Networks (CNNs) have demonstrated significant promise in image classification tasks across various domains, including agriculture. However, the classification of Tikog leaves a culturally significant raw material used in the banig weaving industry in the Philippines has not been explored using CNNs with feature engineering. This study developed and optimized a feature-engineered CNN model for Tikog leaf classification by integrating Lab color space representation, data augmentation, autoencoder-based feature extraction, mean-max pooling, and dropout regularization. A total sample size of 500 standard-quality and 500 substandard-quality Tikog leaf images was augmented to generate 3,000 training images and 500 validation samples. Among the 27 CNN configurations tested, four models demonstrated superior performance, with Case 12 emerging as the best. This model achieved training and validation accuracies of 94.23% and 96.83%, F1-scores of 94.35% and 96.87%, ROC/AUC scores of 98.18% and 99.40%, and low sum of squared errors (SSE) values (173, 19). Case 12 exhibited excellent generalizability, high classification performance, and computational efficiency, making it the most effective model for deployment in real-world Tikog quality assessment. The study advances both technological innovation and the preservation of indigenous knowledge through intelligent systems.

Autoencoder-based Feature Extraction Leverage Artificial Intelligence Image Processing Deep Learning Neural Network Classification Color Representation Mean-Max Pooling Dropout CNN Cultural Preservation