Development of elements of a digital system for identifying soil lumps and separating them from potato tubers in a potato harvester

Equipping mechanical cleaning devices with automation elements does not fully ensure the quality of cleaning commercial products from mechanical impurities. A possible solution is a system of digital identification of separated biological objects from mechanical impurities based on hyperspectral reflection. The authors carried out research to study approaches to processing hyperspectral reflection data of potato tubers and soil clods using the model of neural network YOLOv8 digital system of identification of soil clods mixed with potato tubers in a potato harvester. They have developed a structural scheme of a potato harvester with a digital system of identification of biological objects based on the hyperspectral reflection of potato tubers and recognition of soil clods in an automated mode. Potato tubers and soil clods in the images were recognized using trained models with the selection of lesion areas. The authors obtained curves of accuracy and completeness of soil clods recognition in the wavelength range between 500 and 700 nm. The optimal confidence level for neural network models was 0.28 for soil clods and 0.37 for potato tubers. The authors determined the calculation parameters of binary and multiclass classification metrics of the developed convolutional neural network models at different wavelength ranges for the classes “Potatoes”, “Soil on potatoes”, and “Soil clods”. The authors found that the highest average recognition accuracy for soil clods of mAP 0.329 and that for potato tubers of mAP 0.407 was recorded at a wavelength of 600 nm. The hyperspectral data obtained can significantly increase the accuracy of classification and recognition of diseases and lesions on tubers and can be used to identify minor changes in tuber condition. The accuracy of non-invasive recognition of soil clods and potato tubers using hyperspectral images is comparable to that of human experts (with a deviation of no more than 11.3%).