1D-inception-resnet for NIR quantitative analysis and its transferability between different spectrometers

The development of portable near-infrared spectrometer has greatly enriched the Near Infrared Spectroscopy(NIRS) applications. However, the well-built NIR quantitative model has high requirements for spectral homology, which are not applicable to new spectra acquired by different spectrometers, and model transfer is essential in the field of NIRS technology. Deep learning has become one of the important approaches in chemometrics. For the NIRS datasets of sugarcane samples that acquired by three types of spectrometers, we proposed a deep transfer learning methodology based on an improved one-dimensional Inception Resnet (1D-Inception-Resnet) neural network to establish the quantitative models of two chemical compositions in sugarcane and demonstrate its transferability between spectrometers. First, we developed the primary 1D-Inception-Resnet networks for ADF and IVOMD compositions and explained the distribution differences of the spectral features of the homologous spectra (device 1) and heterologous spectra (device 2 and 3) extracted by the primary networks. Then, we accomplished transferring the primary models to spectrometer 2 and 3 based on deep transfer learning method of fine-tuning the weights of convolution layers, we demonstrated the basic mechanism of deep neural network transfer learning by visualizing the adjustments of spectral features and adjustments of CDF curves obtained by “fine-tuning”. Finally, we compared with traditional model transfer methods. The results prove that the primary 1D-Inception-Resnet quantitative model outperformed PLS primary with exceeding accuracy, combing with fine-tuning transfer learning method, the transferred models have achieved excellent and stable prediction ability on new spectrometers. The “Fine-tuning-1D-Inception-Resnet enabled NIR technique” in this study provides a new idea to promote the practical applications of rapid NIRS detection technology.