Autonomous agricultural machines offer a promising solution to agri-food labour shortages; however, these machines face additional challenges when navigating off-road, often encountering rough terrain which can damage vehicles and reduce autonomous navigation accuracy. Unfortunately, adoption remains limited due to concerns about safety, liability, cost, and lack of trust amongst end-users. This thesis investigated the use of interpretable machine learning for the autonomous perception tasks of terrain roughness prediction and classification, fatigue damage estimation, and obstacle detection Terrain roughness was quantified using the ruggedness coefficient, defined as the ratio of vertical acceleration to vehicle speed. For each model, the input feature set included LiDAR-derived terrain features and vehicle kinematics. A Decision Tree classifier was trained to predict whether ruggedness coefficients exceeded a threshold associated with increased risk of adverse effects. It achieved 99.6% accuracy using only two features (vehicle speed and maximum point-to-plane distance of a plane encompassing the tractor). Using these two features, a Gradient Boosting regressor was trained to predict ruggedness coefficients (R^2 = 0.88). These models enable terrain roughness estimation prior to traversal, supporting path optimization to avoid high-magnitude events which may cause damage or disrupt navigation. Relationships between terrain characteristics, vehicle kinematics, and fatigue damage were also investigated. Fatigue was quantified at the chassis using the Rupp Filter and the Durability Transfer Concept. While the machine learning models did not successfully predict fatigue damage, analysis indicated that most damage occurred in the 0-2Hz frequency band. This coincides with the known frequency range of terrain-induced excitation. Surprisingly, the greatest accumulation of damage occurred during on-road driving, likely due to increased tractor speeds and harder driving surfaces. Finally, a Decision Tree model for real-time obstacle detection was developed. Reusing features from the ruggedness coefficient task, this model achieved an accuracy of 74% and an evaluation time of 23ms, outperforming uninterpretable models identified in literature. This research demonstrates that interpretable machine learning models which leverage LiDAR and vehicle dynamics can enable accurate, efficient, and transparent perception models for off-road autonomous vehicles, supporting greater trust and technology adoption amongst farmers.