外文摘要：Monitoring coffee fruits maturity is pivotal in the decision-making process, particularly during harvesting. However, the coffee plant produces fruits of different maturity stages due to uneven flowering. Previous studies have focused on post-harvesting or reconstruction techniques to monitor coffee crops, but without a significant impact on the decision-making of management strategies, as it needs to occur before harvesting using a scalable and accurate system. Our objective in this study was: (i) to modify and employ a state-of-the-art object detection model and use it to detect and classify coffee fruits based on their maturity stages, enabling rapid and noninvasive monitoring of coffee plants and (ii) to address challenges with image-based detection in coffee field conditions. Therefore, we analyzed pre-harvesting conditions to detect coffee fruits and classify their maturity stage using a YOLOv8 model enhanced through a modification in the convolution block (RFCAConv) to increase performance without compromise computational resource. We also compared image acquisition under two illumination (natural and artificial) and three acquisition conditions: the entire third of the plant (upper, middle and bottom parts), individual branches within the plant, and individual branches against a controlled background. The proposed model achieved mAP@0.50 of 74.20 % and outperformed the standard version YOLOv8n by 1.90 % with minimal increase in computational resource. The AP for unripe, semi-ripe, ripe and overripe fruits were respectively 73.40 %, 67.10 %, 74.40 % and 71.90 %. When comparing the acquisition settings, the highest mAP was obtained when capturing images from branches against a controlled background under natural illumination. (mAP@0.50 of 72.70 %). However, branch and plant-level images also obtained relevant performance and benefit from artificial illumination. Therefore, our study presents a timely contribution as it enables the monitoring of coffee maturity before harvesting through an enhanced lightweight, state-of-the-art detection model, facilitating decision-making. Moreover, our insights further advance the progress of intelligent harvesting systems by addressing diverse field conditions and challenges for the detection and classification of coffee fruits.
外文关键词：object detection;Precision farming;Agriculture 4.0;Coffea arabica
作者：Carreira, Vinicius Dos Santos;Tedesco, Danilo;de Oliveira, Mailson Freire;Kazama, Elizabeth Haruna;Barbosa Jr, Marcelo Rodrigues;Ferreira, Francielle Morelli;Maldonado Jr, Walter;Silva, Rouverson Pereira da
作者单位：Kansas State Univ;Auburn Univ;Sao Paulo State Univ;Mato Grosso State Univ
期刊名称：SCIENTIA HORTICULTURAE
期刊影响因子：0.0
出版年份：2024
出版刊次：328
原文传递申请：江苏省科技资源（工程技术文献）统筹服务平台