Fruit yield estimation using image analysis is also about correcting the number of detections

The use of image analysis to estimate fruit tree yields is increasing. Tools are often based on RGB images acquired from the two opposite sides of the tree and fruit detection using machine learning algorithms. Due to fruit occlusion and detection errors, a correction model is needed to estimate the actual number of fruits from the number of detections. This study aimed to establish the best correction model for estimating mango yield in a diversity of orchards without using any other data extracted from the image than the number of detections. A set of images was acquired at the mature fruit stage on 325 ‘Kent’ mango trees from 27 orchards representing the main cropping systems (from monoculture to diversified orchards) found in two distinct production regions in West Africa. Mangoes have been detected with a YOLOv5 neural network to obtain the detected count. The actual count of fruits was measured on all trees and used to calibrate different correction models for tree production estimation. Results showed that a linear model with region and cropping system as categorical covariates obtained the best performance (R²=0.66, normalised RMSE or NRMSE=11%) compared to a simple linear model (R²=0.34, NRMSE=20%). Square root transformations of actual and detected counts were necessary to fit the regression assumptions of residuals normality and homoscedasticity. Although the use of linear mixed-effect models with the orchard as a random effect might be interesting, their performances were comparable to linear models and they did not respect the regression assumption. This study showed that one simple linear regression might be insufficient when applying fruit yield estimation in different orchards, even for the one cultivar and the same fruit development stage. We also highlight the need to test the regression assumptions before applying the correction model.