Predicting multiple quality traits of citrus: A multi-task LightGBM approach integrating image features

This study proposed a multi-task learning framework based on LightGBM to predict three citrus quality traits simultaneously: soluble solids content (SSC), titratable acidity (TA), and vitamin C content (VC). A total of 182 citrus samples were collected and four images were taken for each citrus, including one from pedicel, one from navel, and two from equatorial planes. First, 26 image features (color indices and texture metrics) and 5 physical features (LAB color space values, fruit size, and weight) were processed through standardization, feature selection, and derivation of agronomically meaningful indicators (aspect ratio and color saturation). Target variables were preprocessed with Box–Cox transformations to correct non-normal distributions. The multi-task learning strategy markedly enhanced predictive performance: the R 2 values for SSC, TA, and VC reached 0.9576, 0.9516, and 0.9730, respectively, with corresponding root mean square error (RMSE) of 0.3885, 0.0504, and 5.0344. Error-interval analysis showed that more than 98% of predictions fell within the ±5% error range, confirming the reliability of the proposed approach. These findings demonstrate the potential of the multi-task LightGBM framework as a practical tool to enable fast, simultaneous and accurate prediction of SSC, TA, and VC using RGB images and simple physical features in the citrus industry.