Historic agricultural terraces are important cultural and geomorphic features, but their spatial distribution is often poorly documented, especially in abandoned landscapes where woodland expansion obscures terrace morphology. This paper presents a fully open-source GeoAI workflow for semi-automatic terrace mapping from high-resolution LiDAR DEMs. The protocol combines predictors derived from terrain morphology and broader landscape context within a Random Forest classification framework. Two models were compared: a benchmark model based on topographic and geomorphological variables, and an extended model incorporating potential solar irradiance, soil erodibility and least-cost corridor density as proxies for agricultural suitability, slope management and accessibility. The extended model consistently outperformed the benchmark model, achieving higher accuracy, precision, recall, F1-score and class separability, while also reducing relative overprediction and omission error. These results show that terrace detection improves when local morphology is combined with predictors that reflect human land-use choices. The workflow provides a reproducible FOSS approach for mapping both visible terraces and those preserved beneath woodland canopy, not only for supporting heritage documentation but also offering potential insights for land degradation assessment and climate adaptation strategies.