In-situ classification of football sports surfaces: leveraging machine learning for enhanced surface analysis

Abstract Sports turf surfaces, including natural turfgrass and synthetic turf, are complex systems with many parameters influencing their performance. This study aims to classify sports turf surfaces using data collected from a bespoke testing device, fLEX, which calculates seven separate metrics related to sports surface performance in both an acceleration format (designed to simulate an athlete accelerating) and deceleration format (designed to simulate an athlete decelerating). Sixty-eight collegiate and professional sports surfaces across the USA and UK were tested, covering a range of climates and field constructions. Surfaces were classified as cool-season, warm-season, or synthetic turf. After data preprocessing, including outlier removal and imputation, two machine learning models, decision tree and random forest, were trained and tested on the dataset. Feature importance was assessed using mutual information, revealing that recoil distance and maximum vertical force were the most critical variables for classification. The decision tree model achieved an accuracy of 84% for acceleration and 79% for deceleration, while the random forest model performed slightly better, with accuracies of 89% and 83%, respectively. Both models demonstrated low overfitting risk, with a minimal difference between training and testing accuracies. Misclassifications were analysed, highlighting the complexity of surface characteristics and potential for improving classification accuracy. The high performing models suggest that the fLEX testing device is an appropriate tool to classify the surfaces, and that unique characteristics exist within each surface category. Collectively, these findings represent a step toward advancing our understanding of the complexity of sports turf surfaces.