Optimized Deep Learning for Tea Leaf Age and Quality Classification using EGACNN and SHEDA-based Hyperparameter Tuning

Reliable tea leaf quality classification is vital for maintaining product consistency and consumer satisfaction. However, traditional assessment techniques—such as physicochemical analysis and chromatographic methods—are labor-intensive, time-consuming, and unsuitable for high-throughput scenarios. Although deep learning models like CNNs, RNNs, and ensemble-based architectures have been employed for automated classification, they often encounter limitations related to hyperparameter tuning, computational inefficiency, and limited generalization across diverse datasets. The primary objective of this study is to develop a computationally efficient and highly accurate framework for tea leaf quality classification by introducing an Ensemble Genetic Algorithm-based Convolutional Neural Network (EGACNN) model. To address hyperparameter optimization challenges, the model is enhanced using a Surrogate-Assisted Hybrid-Model Estimation of Distribution Algorithm (SHEDA). This integration facilitates intelligent and adaptive tuning of critical parameters—including learning rate, batch size, number of convolutional layers, and filter dimensions—while significantly reducing training time and GPU utilization. The EGACNN architecture employs a strategic ensemble of multiple CNN variants to strengthen feature extraction and improve classification robustness. SHEDA further enhances optimization by approximating the search space, enabling efficient exploration of mixed-variable and large-scale parameter domains. The proposed model is trained and validated on the Age-Stratified Tea Leaf Quality Classification Dataset, and benchmarked against state-of-the-art CNNs, RNNs, ResNet, VGG, and hybrid ensemble approaches. Experimental results confirm that EGACNN-SHEDA achieves a classification accuracy of 98.7%, surpassing existing methods in terms of F1-score, recall, and precision. Additionally, the GA-SHEDA integration ensures superior training efficiency and scalability without compromising predictive performance, making it a robust solution for real-world tea leaf quality assessment.