Semantic Segmentation of Golf Courses for Course Rating Assistance

This paper introduces a system to assist golf course raters in determining the difficulty rating of a golf hole. Currently, determining the rating of a given golf hole relies on time-consuming manual measurements on the ground, which we attempt to partially automate. A U - net neural network is trained to classify greens, fairways, tees, bunkers, and water in golf courses, and a course rating assistance system is implemented to measure distances between relevant course parts. Since no public datasets containing golf courses existed prior to this work, we present a new public data set of golf courses created from orthophotos. 1,123 RGB orthophotos for training/validation and 108 RGB orthophotos for testing were gathered from 107 Danish golf courses (58 % of Danish courses) during the spring season and manually annotated. The dataset is publicly available on Kaggle
^1
1
https://www.kaggle.com/datasets/jacotaco/danish-golf-courses-orthophotos. The U-net model accomplished a mean intersection over union (IoU) of 69.6%, mean sensitivity of 78.0%, and mean positive predictive value (PPV) of 84.1 %. Based on this automatic analysis of the course images, the course rating assistance system computes 5 crucial distances for course raters to determine a course rating and achieved a mean error of 2.7% and 17.7 % for green length and width, as well as a mean error of 3.3 % and 4.2 % for male/female hole lengths.