A Novel and Computationally Efficient Chunk-wise Fully Connected Deep Learning Model for Non-Destructive Protein Prediction Using NIR Spectroscopy

Deep learning has emerged as a powerful approach for predicting biochemical traits from spectral data, providing rapid, high-throughput, and non-destructive alternatives to conventional laboratory analyses. Although architectures such as one-dimensional Convolutional Neural Networks (1D-CNNs) and Transformers have shown strong predictive capability, they often require substantial computational resources and large training datasets. In agricultural applications; however, computational efficiency and model generalizability are equally important. To address this need, the present study developed a chunk-wise fully connected neural network that progressively integrates spectral features in fixed-size segments. The proposed architecture was evaluated alongside standard 1D-CNN and Transformer-based models for seed protein prediction using near-infrared reflectance spectroscopy (NIRS) data from three crops: chickpea (n=795), cowpea (n=472), and rice (n=985). The chunk-wise model achieved predictive accuracy comparable to that of the Transformer model, with R 2 values ranging from 0.90 to 0.97, and outperformed the 1D-CNN in both accuracy and Residual Prediction Deviation (RPD). Relative to the Transformer, it required approximately 78% fewer parameters and nearly fourfold less training time. The chunk-wise network also showed consistent performance across all three crops, indicating its applicability across datasets with differing spectral and compositional characteristics. Overall, the proposed model provides a computationally efficient and scalable solution for NIR-based prediction of biochemical traits, with strong potential for integration into portable screening tools and high-throughput breeding pipelines, particularly for large-scale germplasm evaluation.