Evaluating the impact of hyperparameters on the performance of 1D CNN model for nutritional profiling of underutilized crops using NIRS data

The application of 1D Convolutional Neural Networks (CNNs) for nutritional profiling using Near-Infrared Spectroscopy (NIRS) data has increased significantly in recent times. The accuracy of 1D CNNs depends on hyperparameters, yet a thorough investigation into their impact on model performance is lacking. Therefore, in this study, we explored the effects of hyperparameters on the performance of 1D CNN models for predicting nutritional traits using NIRS data from perilla, lablab bean, and rice bean crops. We tested a range of hyperparameters, including convolutional layers, filters, kernel sizes, pooling methods, batch sizes, activation functions, and optimizers. Our results show that increasing the number of convolutional layers improved the model's predictive power by enhancing feature extraction; however, beyond a certain limit, performance declined due to overfitting. Similarly, increasing the number of filters enhanced performance, but the optimal number needs to be decided to mitigate underfitting. Moderate kernel sizes struck a balance between feature preservation and generalization, while larger kernels decreased accuracy. Max pooling with smaller sizes provided optimal results by retaining essential features, whereas larger pooling sizes caused information loss. Smaller batch sizes were more effective at improving generalization, while larger batch sizes led to over-smoothing. LeakyReLU outperformed standard ReLU by avoiding the problem of dead neurons. And, a lower learning rate combined with the Adam optimizer resulted in smoother convergence and higher accuracy. Thus, the present study provides a framework for selecting hyperparameters in 1D CNN models to achieve optimal performance in nutritional trait estimation using NIRS data.