The authors of the Danish Golf Courses Orthophotos dataset present a system designed to assist golf course raters in evaluating the difficulty rating of golf holes. Traditionally, this process involves time-consuming manual measurements on the golf course, which the authors aim to automate partially. They achieve this by training a U-net neural network to classify various elements on golf courses, including green, fairway, tee, bunker, and water bodies. This system helps measure distances between relevant course components. Notably, prior to this work, there were no publicly available datasets containing golf course data. To address this gap, the authors introduce a new public dataset of golf courses created from orthophotos. This dataset comprises 1,123 RGB orthophotos for training and validation, along with 108 RGB orthophotos for testing, all collected from 107 Danish golf courses during the spring season. The dataset includes manual annotations.

In the context of golf, a handicap system allows players of varying skill levels to compete fairly. This system relies on players’ skill levels, known as handicap indexes, and the course rating to allocate the number of shots players receive on each hole based on their playing handicaps.

Historically, determining the rating of golf courses has required labor-intensive measurements performed by a team of four golf course raters. These measurements encompass various factors, including fairway width, green size, distances to obstacles, bunker depth, grass length, tree presence, green slope, elevation, and topography. These measurements are then used to compute the course rating using a complex and confidential set of rules. This manual process is both time-consuming and limits measurement precision and consistency. The authors highlight the advances in satellite observation systems in the 20th and 21st centuries, which have improved the acquisition of accurate remote sensing data for land cover analysis and green space monitoring. They propose that some metrics relevant to golf course ratings could be extracted from overhead imagery.

For data collection, the authors utilized the national database “Dataforsyningen” (DF) to obtain 1,123 RGB orthophotos from 107 Danish golf course during the spring season. These images were randomly divided into training and validation sets in a 70/30 split. Expert annotations focused on classes such as greens, fairways, tees, bunkers, water, and background. Trees and classes that couldn’t be reliably differentiated at a scale of 1:1000 were omitted. An additional 108 images were collected for testing, spanning scales of 1:1000, 1:1250, and 1:1500 to ensure that testing images represented the layout of a single full golf hole. The surroundings of the golf hole were manually masked in the testing images to facilitate course rating computations per hole.