Abstract
AbstractNear infrared spectroscopic (NIRS) data from different harvested seasons may consist of different feature spaces even though the samples have the same label values. This is because the spectral response could be affected by the changes in environmental parameters, internal quality, and the reproducibility of NIRS instruments. Thus, this study aims to investigate the ability of Joint Distribution Adaptation (JDA) transfer learning algorithm in addressing the assumption of traditional machine learning i.e. both training and testing data must come from the same feature spaces and data distribution. First, NIRS data acquired from two different harvested seasons were used as the source domain and the target domain, respectively. Next, JDA was implemented to produce an adaptation matrix using the source domain and transfer datasets. This adaptation matrix would be used to transform the source and target domain datasets. After that, a calibration model was developed by means of Partial Least Squares (PLS) using the transformed training dataset; and validated using the transformed independent testing dataset. The proposed JDA-PLS was compared to the PLS without transfer learning as the baseline learning. Findings show that the proposed JDA-PLS with 10 LVs achieved the lowest RMSEP of 1.134% and the highest RP2 of 0.826.KeywordsTransfer learningJoint distribution adaptationNear-infrared spectroscopyDifferent seasons
Accepted Version (
Free)
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call