Comparative evaluation of traditional and novel neural Network-Based approaches for spectral transfer in MEMS FT-NIR...

Fourier Transform Near-Infrared (FT-NIR) spectroscopy faces critical challenges in calibration trans- fer, particularly in achieving consistency across diverse instruments and environmental conditions. This study evaluates traditional methods-Spectral Space Transformation (SST) and Spectral Align- ment-alongside a novel Spectral Residual Neural Transfer Network (SRNTN) on Micro Electro Mechanical System (MEMS)-based FT-NIR sensors. A meticulously curated standard calibration set, consisting of 28 diverse samples, including agricultural feed materials and inorganic substances, was used as the foundation for spectral transfer. Data collection was conducted using 8 spectral sensors, with two of the samples being stable salt samples containing moisture: SrCl₂·6H₂O and SrCl₂·6H₂O + 10 % TiO₂. Using this dataset, along with supplementary dry-ground, dry-whole, and wet-whole materials, we assess performance based on Root Mean Squared Error (RMSE), correlation coefficient squared (R²), and mean range improvements. SST methods demonstrated robust performance for dry and ground materials, achieving up to a 45.9% improvement in mean range and a 26.9% reduction in RMSE. Spectral Alignment proved highly effective for wet and whole samples, with mean range improvements reaching 18.8%. The proposed SRNTN method exhibited notable adaptability and consistency, achieving mean range improvements of up to 38.4% and RMSE reductions of 21.9% across specific datasets. Additionally, SRNTN demonstrated the least variability in mean range improvement percentages across all datasets, with the lowest standard deviation in this metric, and achieved the best minimum and maximum improvements among all methods, underscoring its robustness and reliability. Furthermore, a novel moisture superposition approach significantly enhanced performance for wet samples, achieving up to a 17.3% RMSE improvement-nearly three times the best improvement observed prior to superposition. These findings highlight the critical role of tailored calibration strategies in ensuring reliable and accurate spectral transfer.