A CNN-based approach to detect cover damage of round cotton modules

Authors: IQBAL, M - Texas A&M University; HARDIN, R - Texas A&M University; WARD, J - North Carolina State University; Wanjura, John

Submitted to: National Cotton Council Beltwide Cotton Conference
Publication Type: Proceedings
Publication Acceptance Date: 2/26/2022
Publication Date: 4/1/2022
Citation: Iqbal, M.Z., Hardin, R.G., Ward, J.K., Wanjura, J.D. 2022. A CNN-based approach to detect cover damage of round cotton modules. National Cotton Council Beltwide Cotton Conference. Memphis, TN. 616-623.

Interpretive Summary: The adoption of cotton harvesters that form round cotton modules during harvest are gaining wide spread adoption across the US cotton belt due to their ability to reduce harvest-time labor requirements and enhance transportation logistics. However, the plastic used to wrap and protect the modules is a source of contamination in ginned lint due in part to the inclusion of plastic from wraps damaged during handling and transport. It is estimated that plastic contamination cost the US cotton industry about $750 million per year. To help address contamination issues resulting from damaged module wraps by improving the ability for gins to implement contamination prevention strategies, an automated image collection system was developed and was used to capture images of modules during handling with a wheel loader. To automate the identification of damaged module wraps, three computer models were evaluated to identify the best performer. The YOLOv5model performed best, detecting cover damage of the round cotton modules with an accuracy of about 75%. The detection accuracy of the model was not as high as desired. Future work to improve detection accuracy through additional model training and improved image processing techniques to remove background effects is planned.

Technical Abstract: Round cotton modules covered with engineered plastic film are increasingly popular because they largely automate the cotton harvest. Cover damage may occur during handling of the round cotton modules, resulting in plastic contamination in the processed cotton fiber. Plastic contamination costs millions of dollars every year to the US cotton industry. To uphold its reputation for providing some of the cleanest cotton to the international market, the US cotton industry is prioritizing the removal of plastic contamination from cotton. To produce contamination-free fiber, it is important to determine detailed information about plastic cover damage of cotton modules. Farmers and ginners need to identify damaged covers to be able to repair or handle carefully; consequently, an automatic cover damage identification system would be useful. Therefore, the objective of this research was to use different convolutional neural network (CNN) models to classify the cotton modules with damaged covers during the handling process and identify the best-performing one. A single-board computer-based system was developed and installed on a loader to collect images of the cotton modules during the handling process. This system collected images of the cotton modules from different directions during the handling process and stored them with a unique RFID number. The YOLOv5 CNN model was most accurate in detecting cover damage, with training, validation, and testing accuracy of 75.6%, 65.9%, and 78%, respectively. This model was able to provide information about the status of each module from the images collected by the system with 73% accuracy. The model developed in this study might be a useful tool to reduce plastic contamination in processed cotton.