A Deep Learning Approach to Detect Severity of Mango Damage in the Early Fruit Stage

Accurate detection of mango fruit damage predominantly from fruit fly infestation is pivotal as it directly affects both yield and trade worldwide. Therefore, timely identification of such damage is critical to mitigating the spread of infestation and minimizing associated losses. This paper focuses on the early detection of mango damage in orchards using YOLOv8 models, that offer enhanced accuracy and speed compared to earlier versions, making them more efficient for object detection tasks. Limited studies have been done to detect and classify damage on fruits in orchards using deep learning, with the need for more models to detect various categories of damage instances. The experiments in this study revealed no substantial differences among the various YOLOv8 versions used with the highest accuracy of 88.6% and 98.5% attained for detecting damage and mango instances respectively. Both YOLOv8s and YOLOv8l obtained a precision value of 88.6% for lesion detection, and 87.9% using YOLOv8x. However, YOLOv8x achieved slightly higher values of recall and mAP compared to other models in detecting damage features. The study has further revealed that learning the damaged features of mango fruit is more challenging compared to healthy features, as observed from values obtained from the precision-recall curve. Through fine-tuning parameters of the models, our experimental results using the YOLOv8 model demonstrate the potential of lesion detection on mango fruits on trees, leveraging a dataset of 1317 images augmented to 3161. This study addresses the challenge of estimating profits and losses for fruits still on trees, which has been relatively overlooked in prior research efforts. We believe this method can effectively be adapted to detecting lesions on other fruits in orchards with minimal modifications. Future work can consider a better dataset with minimal noise while exploring different growth stages of fruits, and weather conditions the data is captured using alternative models while incorporating other factors in the segmentation and analysis phases.