Transfer of apple soluble solids content prediction model across cultivars based on domain-adversarial neural network

Abstract

Poor model generalizability has become a key problem in using visible and near-infrared (Vis-NIR) spectroscopy to detect the internal quality of fruit, which is due to the differences in the physicochemical properties and spectral distributions of fruit under different cultivars, leading to the fact that the model established for a certain cultivar cannot effectively predict the internal quality under other cultivars. To solve this problem, 'Ruixue', Luochuan 'Fuji', and Jingning 'Fuji' were used as samples to compare the transfer performances of three transfer methods, namely, two-stage TrAdaBoost.R2 (TTB), fine-tune (FT), and domain-adversarial neural network (DANN). The results showed that the model based on DANN could effectively eliminate the variability among different cultivars in the same region and year, which was more effective in transferring the prediction model of soluble solids content (SSC). In the transfer of 'Ruixue' to Luochuan 'Fuji', when samples of the target domain were 67, the $R_p$ was 0.93, and the $\text{RMSEP}$ was 0.56 %, confirming the potential of DANN in solving poor model generalization.