Accurate prediction of soluble solid content of apples from multiple geographical regions by combining deep learning with spectral fingerprint features

摘要

The geographical origin of an apple can affect its cellular structure, and therefore its optical properties including interactions with incident light. As a result, accurate prediction of soluble solid content (SSC) in apples with multiple geographical origins is still challenging. A multiple-origin SSC prediction model for apples from multiple geographical regions has been developed by combining spectral fingerprint feature extraction, origin recognition, model search strategies, optimal wavelength selection, and deep learning with multivariate regression analysis. In this model, the spectral fingerprint features of apples were explored and determined using the random frog algorithm, and deep learning was used to train and test for origin recognition with the fingerprint spectral feature as inputs. Particle least squares (PLS) was applied to develop individual-origin calibration models, and subsequently employed to detect SSCs. A competitive adaptive reweighted sampling (CARS) algorithm was used to select the optimal wavelengths for the calibration models. Compared with the individual-origin model, the proposed multiple-origin model achieved more accurate results for the prediction of SSC of apples with multiple geographical origins, with the R-P and RMSEP values being 0.990 and 0.274, respectively. These results indicate that variations in geographical origin affect accuracy, but that the multiple-origin model can eliminate the effects of geographical origin on SSC prediction, thereby improving the applicability of SSC detection in practice.