Monitoring coffee fruit maturity using an enhanced convolutional neural network under different image acquisition settings

Abstract

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 non-invasive 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.