Predicting Artificial Turf Stiffness using Wearable Sensors and Machine Learning

Abstract

Artificial turf stiffness affects performance and injury risk, but current testing devices are costly, inefficient, and limited in coverage. This thesis developed and validated a wearable sensor and machine learning system to predict turf stiffness as a practical alternative. Five turf surfaces with distinct stiffness levels were tested with 50 athletes performing walking, running, and jumping. Shoe-mounted IMUs and pressure insoles captured biomechanical data. Machine learning models, including an ensemble of Support Vector Machine, Random Forest, and Gradient Boosting, as well as a 1D-CNN, were evaluated with Leave-One-Subject-Out cross-validation and validated on a professional field. The walking movement proved to contain the most valuable information for prediction performance and was used solely for final modelling. The ensemble achieved the highest performance, classifying five stiffness levels with 74.0 ± 9.1% accuracy and 99.8 ± 0.1% adjacent class accuracy. Regression models predicted continuous stiffness values with good fit (R² = 0.836 ± 0.085). The 1D-CNN reached 64 ± 5.1% accuracy, showing moderate generalization compared to feature-based models. Generalization testing on McMahon Stadium’s professional field confirmed strong agreement with gold-standard devices, with 96.5% of trials correctly classified and regression outputs closely matching mechanical measurements. Athlete surveys revealed inconsistent perception of stiffness, highlighting the need for objective monitoring. Wearable-based machine learning enables accurate, scalable, and cost-effective prediction of turf stiffness, providing a practical tool for real-time monitoring and field management