Tree-level citrus yield prediction utilizing ground and aerial machine vision and machine learning

Yield prediction of citrus provides critical information before harvest to growers and allied industry to predict the resources required for workers, storage, and transportation of the harvest. In this study, three machine learning (ML) based models were developed for tree-level citrus yield prediction: (i) Model-1 utilized UAV imagery; (ii) Model-2 utilized UAV imagery and ground-based fruit detection and counts from images taken from one side of the tree; and (iii) Model-3 utilized UAV imagery and ground-based fruit detection and counts from images taken from two sides of the tree. The UAV images were used as input to a novel cloud-based technology, Agroview, to get the tree health, height, and canopy area information. The multispectral bands and the tree structural parameters were the input for Model-1. Two images per tree were captured from the ground using an RGB camera (one from each side) and were used for fruit count using an object detection algorithm, YOLOv3. Harvest data was collected manually per tree (fruit count and weight). Four ML algorithms - gradient boosting regression (GBR), random forest regression (RFR), linear regression (LR), and partial least squares regression (PLSR) were used to generate the models. Model-2 (MAPE of 23.45%) performed similarly to Model-3 (MAPE of 25.72%) and significantly better than Model-1 (MAPE of 35.59%). Model-2 was selected as the best model because of its low MAPE value in predicting yield at the tree level, and data collection simplicity (compared to Model-3).