A performance comparison of RGB, NIR, and depth images in immature citrus detection using deep learning algorithms for yield prediction

Yield forecasting is important for farm management. In this study, red, green, and blue (RGB), near-infrared (NIR), and depth sensors were implemented in an outdoor machine vision system to determine the number of immature citrus in tree canopies in a citrus grove. The main objective was to compare the performances of three image data types for citrus yield forecasting. The performance comparison was conducted with two machine vision algorithm steps: 1) circular object detection for potential fruit areas and 2) classification of citrus fruit from the background. For circular object detection, circular Hough transform was used in the RGB and NIR images. For the depth images, CHOI‘s Circle Estimation (‘CHOICE‘) algorithm was developed using depth divergence and vorticity to find circular objects in the depth images. The classification process was conducted using AlexNet, a deep learning algorithm for all three image types. The implementation of a convolutional neural network allowed the machine vision algorithms to remain bias-free process during feature generation and selection. NIR images performed best with 96% true positive rate for both the circular object detection and classification. A machine vision system using this image type will produce a more objective yield prediction with a higher accuracy than other types.