Estimating melon yield for breeding processes by machine-vision processing of UAV images

Monitoring plants, for yield estimation in melon breeding, is a highly labor-intensive task. An algorithmic pipeline for detection and yield estimation of melons from top-view images of a melon's field is presented. The pipeline developed at the individual melon level includes three main stages: melon recognition, feature extraction, and yield estimation. For each region of interest classified as a melon, the melon features were extracted by fitting an ellipse to the melon contour. A regression model that ties the ellipse features to the melon's weight is presented. The modified R2 value of the regression model was 0.94. Comparing yield estimation to ground truth, the average estimation error was 16%. The yield accuracy is highly dependent on the ellipse estimation accuracy, with promising results of only 4% error for the best ellipse-fitted melons.