Automatic detection and classification of agricultural terraces at large scales: a case study of Cyprus

Terraced landscapes are among the most significant anthropogenic modifications in the Mediterranean, essential for agriculture, erosion control, and cultural heritage. Yet, systematic large-scale assessments of their occurrence remain limited. This study presents a reproducible and computationally efficient framework for object-based detection and classification of agricultural terraces using freely available geospatial data. The framework combines image segmentation with two-stage machine learning classification and has been applied to Cyprus, a representative Mediterranean setting. The terrace detection model achieved a Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 0.86, a Precision–Recall Area Under the Curve (PR AUC) of 0.40, and a normalized Matthews Correlation Coefficient (MCC) of 0.69. The classification model, distinguishing broad, narrow, and reforestation terraces, reached a normalized MCC of 0.75. Feature importance analysis showed that spectral, elevation, and land-cover variables were most influential for detection, whereas texture features were key for classification. Applying the model at the island scale yielded a total predicted terraced area of 693 km2 (7.5% of Cyprus), equivalent to nearly half of the country’s agricultural land. The resulting maps provide the first island-wide overview of terrace occurrence and types, offering opportunities to support applications in land management, erosion modelling and geoarchaeology.