Quick Count of Different Orange Cultivar on Trees based on Smartphone Image and Artificial Intelligent

Abstract

Currently, predictions of orange fruit yield in an orchard are still done manually, namely by sampling manually to count the number of oranges on the tree. This method is not effective and the accuracy of predictions cannot be guaranteed. Automation in the process of counting citrus fruit on trees to predict yield can be done with computer vision using artificial intelligence models for object detection. One of the proposed model solutions that can be used for object detection is by using You Only Look Once (YOLO) architecture. However, the performance of the YOLO model for different varieties of orange trees in Indonesia is not yet known. Therefore, in this research, the development of the YOLOv5 model was carried out to quickly count orange fruit on trees of different varieties including the stages of image capture, image resizing, segmentation, model training with hyperparameters such as batch size and epoch, as well as model evaluation. In this study, the primary image dataset taken consisted of images of orange trees with two different cultivars, namely Pontianak Siamese oranges and Terigas Tangerines which have different characteristics. Then the YOLOv5 model is trained using labeled image data. The YOLOv5 model is trained with variations of hyperparameters and then the results are compared. The best model results in Siam Pontianak have a single label configuration in batch size 4 with parameters Mean Average Precision (mAP50), accuracy, precision, recall, and F1-score which produces a value of 0.88; 0.712; 0.853; 0.822; and 0.8372. Meanwhile, the best model results in Keprok Terigas have a single label configuration in batch size 10 with parameters Mean Average Precision (mAP50), accuracy, precision, recall, and F1-score which produces a value of 0.933; 0.75; 0.913; 0.878; and 0.8951.