Lameness, limping due to pain, is a significant welfare issue for horses. Veterinarians typically evaluate horses on two terrain types (hard and soft, e.g., asphalt and sand) that are known to affect the observed degree of lameness based on the origin/location of the pain. In the past years, whole-body inertial measurement units (IMU)-based gait analysis systems were developed to support diagnostics and monitor locomotion changes over time. Movement direction and gait (walk, trot) are automatically labeled, resulting in smart and easy-to-use systems. However, terrain types are not detected, leading to information loss. In this work, we explored terrain classification tasks with equine IMU data and machine and deep learning. Using the data of 111 horses equipped with IMU sensors (withers, pelvis, front, and hind limbs), we compared different features-based (FT) and time-series-based (TS) classifiers (train-test ratio: 0.7-0.3). In order to reduce the computational costs of the future system, we also evaluated the performance (F1 score) of the classifiers with different sampling frequencies (10 to 200Hz) and different IMU combinations (body and limbs). Our Convolutional Neural Network models accurately classified terrain types with only one IMU placed on the front limb. Downsampling the signals led to similar results, thus enabling real-time applications.