Fruit detection and load estimation of an orange orchard using the YOLO models through simple approaches in different imaging and illumination conditions

Abstract

Fruit load estimation is an essential step toward Precision Agriculture (PA) as it helps growers more accurately predict market planning, worker planning, purchase of appropriate equipment and so on. Reliable and accurate estimation of fruit yield in an orchard with hundreds of trees needs automatic methods. In recent years, Deep Learning (DL) has been studied widely and applied in various fields of agriculture. Accordingly, the YOLO detection models were applied to detect and count ripe Dezful native orange in an orchard in southwestern Iran. The models were adapted through transfer learning and trained by Google Colaboratory in the RGB images to detect and count orange fruits. Models performance and accuracy of yield estimation for an orchard with 1115 trees were examined. The process was conducted in 3 steps, including training and testing the different versions of the YOLO models by creating an image dataset of orange trees in different illumination conditions, evaluating the models on 100 sample trees, and finally extracting the yield variation map of the orchard after detecting and counting the oranges on images taken from all the trees in the orchard. The precision, recall, F1-score and mAP of the YOLO-V4 as the best model for orange detection over the test images were 91.23%, 92.8%, 92%, and 90.8%, respectively. The overall performance of the models in nighttime and daytime imaging was not significantly different. The YOLO-V4 model was chosen to use for yield estimation in the orchard. The promising results show that the YOLO models can effectively provide researchers and agricultural activists with a simple and practical method for detecting and estimating the yield of orange fruits in an orchard. Significant differences were observed in yield estimation for two-side and four-side imaging. Accordingly, a combined imaging method including two-side and four-side imaging was proposed for thin and dense canopy, respectively. The map of fruit yield changes showed the spatial distribution of tree yield with a +9.19% error.