Deep-Learning-Based Monitoring of Impurity Content and Breakage Rate in Rice Combine Harvesters

Continuous monitoring of impurity content and breakage rate in combine harvester grain flow remains challenging because representative samples are difficult to acquire online, and the visual targets are small, dense, and imbalanced. In this study, a prototype monitoring system integrating sample collection, controlled conveying, image acquisition, and embedded processing was developed for online grain-quality sensing during harvesting. To satisfy the requirement for sidewall sampling from the vertical grain conveying auger, centrifugal sampling and screw conveying were used to extract and transport grain-flow samples, and a stable imaging environment was established using an industrial camera and dedicated illumination. Pixel-area-to-mass mapping models were established for broken grains and impurity targets, with coefficients of determination higher than 0.93. In addition, a lightweight improved YOLOv8-Seg model was developed to recognize and segment broken grains and impurity targets under dense small-target conditions. Bench-scale validation showed that the relative error of impurity content ranged from 1.02% to 13.04%, with an average of 6.09%, while the absolute error of breakage rate ranged from 0.01 to 0.02 percentage points. These results demonstrate the feasibility of the proposed method for online estimation of impurity content and breakage rate under bench-scale conditions and provide a basis for future machine integration and field validation.