Nondestructive evaluation and high-throughput edible rate prediction method for durian based on X-ray CT and machine learning

The sarcocarp content and volume are key factors for assessing the quality of durian fruit. Non-destructive and efficient detection techniques can provide critical data support for durian breeding, postharvest technology research, and grading during industrialization processes. This study primarily investigated the potential of X-ray computed tomography (CT) in non-destructive analysis of internal durian traits and high-throughput evaluation of edible rate. The X-ray CT system was used for durian imaging. The U-Net model was utilized to segment tomographic images into background, sarcocarp, pericarp, kernels, and cavity regions. A total of 18 phenotyping traits, such as fruit volume and sarcocarp volume, were automatically calculated. To overcome the impact of tomographic image quantity on detection efficiency, a rapid prediction model for edible rate was developed using a single image. Experimental results indicated that the method based on X-ray CT system achieved mean absolute percentage error (MAPE) of 1.14 % for fruit volume and 3.09 % for sarcocarp volume when compared to manual measurements, with coefficient of determination (R²) values of 0.989 and 0.955, respectively. When predicting the edible rate based on all tomographic images, the R² and MAPE were 0.923 and 3.39 %. Further results indicated that the edible rate prediction model based on a single image achieved R² and MAPE values of 0.917 and 4.03 %. Overall, X-ray CT imaging technology facilitates comprehensive and accurate extraction of internal durian traits while also demonstrating its potential for high-throughput prediction of the edible rate.