Deep convolutional networks based on lightweight YOLOv8 to detect and estimate peanut losses from images in post-harvesting environments

Abstract

Peanut losses detection is key to monitor operational quality during mechanical harvesting. Current manual assessments faces practical limitations in the field, as they tend to be exhaustive, time-consuming, and susceptible to errors, especially after long work periods. Therefore, the main objective of this study was to develop an automated image processing framework to detect, count, and estimate peanut pod losses during the harvesting operation. We proposed a robust approach encompassing different environmental conditions and training detection algorithms, specifically based on lightweight YOLOv8 architecture, with images acquired with a mobile smartphone at six different times of the day (10 a.m., 11 a.m., 1 p.m., 2 p.m., 3 p.m., and 4 p.m.). The experimental results showed that detecting two-seed peanut pods was more effective than one-seed pods, with higher precision, recall, and mAP50 values. The best results for image acquisition were between 10 a.m. and 2 p.m. The study also compared manual and automated counting methods, revealing that the best scenarios for counting achieved an R² above 0.80. Furthermore, georeferenced maps of peanut losses revealed significant spatial variability, providing critical insights for targeted interventions. These findings demonstrate the potential to enhance mechanized harvesting efficiency and lay the groundwork for future integration into fully automated systems. By incorporating this method into harvesting machinery, real-time monitoring and accurate loss quantification can be achieved, substantially reducing the need for labor-intensive manual assessments.