Back Propagation Neural Network model for analysis of hyperspectral images to predict apple firmness

The potential of employing hyperspectral imaging (HSI) in the near-infrared (NIR) range (386.82−1,004.50 nm) for predicting the firmness of 'Fuji' apples cultivated in Aksu has been evaluated. The performance of seven preprocessing algorithms and two feature selection algorithms was evaluated. The coefficient of determination (R²) and root mean square error (RMSE) of Partial Least Squares (PLS) models are contrasted using various inputs. These results confirm that the Multiplicative Scatter Correction (MSC) preprocessing algorithm was the optimal choice (R²p = 0.7925, RMSEP = 0.6537), and the Competitive Adaptive Reweighted Sampling (CARS) feature selection algorithm demonstrated superior performance (R²p = 0.8325, RMSEP = 0.6257). Based on the aforementioned findings, PLS, Multiple Linear Regression (MLR), Heterogeneous Transfer Learning (HTL), and Back Propagation Neural Network (BPNN) models were constructed for cross-validation purposes. The experimental results indicate that the CARS-BPNN model exhibits the optimal prediction performance, with an R²p value of 0.9350 and an RMSEP value of 0.4654. The results of the research indicated that a deep learning method combined with hyperspectral imaging technology could be utilized to non-destructively detect the firmness of 'Fuji' apples, which will be beneficial and potentially applicable for post-harvest fruit firmness monitoring. This research provides a reference point for the non-destructive detection of apple in the selection of preprocessing, feature selection algorithms, and predicting firmness model.