Prediction of apple leaf moisture content based on color and texture features

Objective To establish an accurate non-destructive prediction method for apple leaf moisture content.

Methods In this study, an industrial camera was used to collect RGB images of apple leaves. Color features were extracted from multiple color spaces, and texture features were extracted by the Grey-level Co-occurrence Matrix (GLCM) technology. Statistical analysis was performed to screen out 34 image features significantly correlated with leaf moisture content. Principal Component Analysis (PCA) was adopted to reduce the dimensionality of the above correlated features for higher model efficiency and prediction accuracy. On the basis of dimension-reduced features, Partial Least Squares Regression (PLSR), Random Forest (RF) and Convolutional Neural Network (CNN) models were constructed respectively.

Results For the three models built with dimension-reduced data, the coefficients of determination of prediction set (R²_P) were 0.62, 0.879 and 0.716; the Root Mean Square Error of Prediction (R_MSEP) were 2.037, 1.102 and 1.954; the Residual Prediction Deviation (R_PD) were 1.636, 3.421 and 1.948 in sequence.

Conclusion This study provides a novel method for rapid non-destructive detection of apple leaf moisture content, and the Random Forest model achieves the optimal prediction accuracy.