CNN-based methods vs. KNN-weighted PLSR: a comparative analysis for protein and adf prediction across multi-product feed.

While traditional chemical analysis methods are precise, they are time-consuming and costly, necessitating rapid, non-invasive alternatives such as Near-Infrared Spectroscopy (NIRS). Deep learning models, particularly Convolutional Neural Networks (CNNs), have demonstrated strong predictive capabilities by effectively extracting spectral features from large datasets. However, their performance compared to local and global calibration methods in multiproduct feed analysis remains underexplored. This study evaluates the predictive accuracy of three CNN-based models—CNN, CNN + Partial Least Squares (CNN + PLS), and CNN + XGBoost—against local calibration approaches, including kNN-Weighted PLSR (kNN-LWPLSR) and PLS methods with Global Calibration Models (GCM). Using a dataset of 3,143 samples from seven different feed products, models were developed using the entire dataset encompassing all sample types, rather than a single feed type. The results in validation (sample n=number) showed that for Protein prediction, kNN-LWPLSR achieved the lowest RMSEP (0.57) and highest RSQ (0.99), while CNN-based models (CNN: 0.62, CNN + PLS: 0.60, CNN + XGBoost: 0.60 with same RSQ: 0.98) demonstrated similar accuracy, outperforming global PLS calibration methods such as GCM (RMSEP: 0.73). For ADF prediction, kNN-LWPLSR achieved the lowest RMSEP (1.03), while CNN-based models (CNN: 1.38, CNN + PLS: 1.39, CNN + XGBoost: 1.62) exhibited higher errors. The global calibration model GCM (1.44) performed similarly to CNN-based approaches but was less effective than KNN, reaffirming the continued relevance of local calibration techniques for fiber estimation. As this study was conducted on the entire dataset covering all feed types not for the single type, the results confirm the generalizability of CNN-based approaches for protein prediction in multi-product feed, while reinforcing the effectiveness of local calibration for fiber analysis.