Spectral prediction of apple fruit quality using samples representing the on-tree variability.

We studied the spectral prediction on apples selected using a multilevel systematic sampling design applied to a group of trees. VIS-NIR spectra in the 400-1130 nm range were collected using a bench-top spectrometer. Soluble solids content (SSC) and firmness were measured using conventional, destructive techniques in the laboratory. Firmness values exhibited higher variability than sugar. The data sets were analyzed using Principal Least Squares (PLS) to calculate prediction models for each for SSC and firmness. Three different ways to select calibration and validation sets from the sample data were compared. The ‘smooth fractionator’ method, resulted in considerably reduced validation-model Bias. With a relatively small number of fruit samples our calibration models results for firmness were r=0.82 and RMSECV=6.15(N) and for soluble solids content r= 0.74 and RMSECV = 2.99 (Brix). The validation results for the prediction models were r=0.59 and RMSEP= 8.69 (N) for firmness and r=0.56 and RMSEP= 1.78 (Brix) for soluble solids content (SSC).