Data fusion of FT-NIR spectroscopy and Vis/NIR hyperspectral imaging to predict quality parameters of yellow flesh “Jintao” kiwifruit

Abstract

The internal quality of kiwifruit, in terms of soluble solid content (SSC), flesh firmness (FF), and dry matter (DM), is widely recognised as a key feature for fruit sorting and pre-harvest assessment. Furthermore, flesh hue (FH) is another important parameter to consider for yellow flesh kiwifruits. NIR and VIS/NIR spectroscopic techniques are valuable alternatives for rapid and non-destructively prediction of all these quality parameters in fruit. Accordingly, the aim of this work was to build a partial least square (PLS) regression models to estimate SSC, FF, FH, and DM of yellow fleshed Actinida chinensis (Jintao) starting from Vis/NIR hyperspectral imaging (400–1000 nm) and FT-NIR (800–2500 nm) spectroscopy data. To take advantage of the complementary information of the two different spectral ranges, data fusion strategies were investigated to concatenate the data before PLS models. In particular, two different sequential fusion methods were used: low-level data fusion based on the concatenation of the pretreated spectra, and mid-level feature fusion characterised by the concatenation of features (scores) obtained by principal component analysis (PCA) or PLS models developed considering individually each data set. For all quality parameters, the best results were achieved by adopting the second approach of mid-level data fusion (PLS scores), reporting RP2 (test set validation) of 0.914 (RMSEP=0.97°Brix), 0.843 (RMEP=1.82°H), 0.866 (RMSEP=9.41N), and 0.854 (RMSEP = 0.64%) for SSC, FH, FF, and DM, respectively. Furthermore, with respect to the PLS models from the individual data sets, the results reported a mean RMSEP reduction of 16.0 ± 4.8%, confirming the potential of the data fusion in improving the PLS prediction power for the quality parameter of kiwifruit.