Spectral Data Classification By One-Dimensional Convolutional Neural Networks

Despite the dominance of Chemometric methods and traditional machine learning algorithms in the field of one-dimensional (1D) spectral data analysis for decades, they still rely on trivial pre-processing steps and hand-crafted feature selections based on domain expertise. Deep two-dimensional convolutional neural networks (2D CNNs) which utilize automatic feature extraction for maximum accuracy have recently achieved overwhelming success in areas such as computer vision and natural language processing. Nevertheless, the 2D CNNs can not be directly applied to the 1D spectral data classification tasks restrained by both the lower dimensionality and scarce samples. Consequently, several dedicated adaptations must be made to ensure the successful implementation of 1D CNN. In this study, a shallow 1D CNN, aiming to attain a compromise between high performance and ease of implementation, was presented. Then, an effective strategy to select architecture hyperparameters and facilitate the fine-tuning process was proposed. Finally, a real case of spectral classification of corn seed viability using visible near-infrared (Vis-NIR) data was carried out, and the 1D CNN achieved 93.3% accuracy, which was superior to the PLS-DA (91.1%) and SVM (86.7%). The results from this study demonstrate that 1D CNN with a highly compact architecture can be devised to effectively classify spectral data in domain-specific tasks.