Running surfaces influence energy consumption and gait parameters including swing time and stance time. This paper compares running gait cycle time, swing time and stance time, recorded on an athletics track, soft sand, and hard sand. The training and evaluation of supervised machine learning models for running surface prediction were developed using an ankle-worn inertial sensor. Models were trained using statistical features extracted from six participants using gyroscope-based stride cycles. Six different model types were trained and the performance of each model was evaluated using precision, recall, F1-score, Matthews correlation coefficient, area under the precision–recall curve and accuracy. There was a significant statistical difference in swing time and stance time across the surfaces for all participants (p < 0.05). Athlete-independent models demonstrated acceptable ability to distinguish soft sand from the two harder surfaces (≥ 0.75 mean precision, ≥ 0.90 mean recall, ≥ 0.83 mean F1-score, ≥ 0.98 mean area under the precision–recall curve across all models), but they were poor at differentiating between athletics track and hard sand. The athlete-dependent models demonstrated strong ability to classify all the surfaces (weighted average precision, recall, F1-score, Matthews correlation coefficient, area under the precision–recall curve, and overall accuracy ≥ 96%). Support vector machine models were the best in both athlete-independent and athlete-dependent methodologies. Features extracted from an ankle-worn inertial sensor can be used to classify running surface with high performance, when models are trained using features pertinent to each athlete.