OBJECTIVE: Expressways in hilly areas feature complex alignment and environments constrained by terrain conditions, significantly threatening life and property safety. This study aims to investigate crash risk prediction of expressways in hilly areas through alignment and environment features and identify determinants of the high risk for safety improvement. METHODS: Based on 5 years of crash data on casualties and property damage of an expressway in southwestern China, the order technique and five clustering algorithms were employed to determine and classify risk levels. Environment features were extracted by semantic segmentation with a DeepLabv3 model. The study established four ensemble learning models to predict crash risks, and the interpretable model approach was adopted to understand contributing features. RESULTS: XGBoost achieved the best overall performance, with the accuracy and F1 score reaching 0.9259 and 0.8886. The proportion and variation rate of trucks and cars, and the proportions of constructions and the road positively correlated with high risks, while the proportions of the vegetation and road had negative correlations. The horizontal and vertical alignments, including long steep slopes, smaller curve radii, shorter transition curves, and smaller convex and concave curves radii, were linked to high risks. CONCLUSIONS: This study proposes an approach to predict crash risks on road sections without historical crash data. Combining the XGBoost model with the SHAP approach, enables accurate identification of risks on expressways in hilly areas using alignment and environment features and enhances the understanding of how these factors contribute to high risks.