Automatic mapping of terrace systems at large scales: a case study of Cyprus

Agricultural terraces are among the most significant anthropogenic land modifications in the Mediterranean. They are constructed to reduce local slope gradients and facilitate farming by artificially increasing soil thickness. Terraces also reduce soil erosion and enable irrigation practices. Yet, if not maintained or abandoned, they become prone to piping, gully erosion, and landsliding. Nevertheless, incorporating these effects into large-scale hydrological, geomorphological, and agronomic research remains challenging due to limited information on terrace locations and characteristics.

We aim to address this gap by presenting a new predominantly automatic approach for detecting and classifying terraced units on a large scale. This scalable tool utilizes freely available data: Google optical satellite imagery (≈2.1 m spatial resolution), ALOS Global DSM (30 m spatial resolution), and ESA WorldCover (10 m spatial resolution). Our study site, Cyprus, with an area of 9,250 km² has a long history of terraced agriculture driven by its rugged terrain. The island features terraces ranging from old, abandoned ones to newly constructed terraces using heavy machinery.

The approach employs Object-Based Image Analysis (OBIA). First, images are segmented using SLIC (Simple Linear Iterative Clustering). These segments are populated with information from 22 derivative layers generated from the initial data. For each segment, 36 statistical parameters are calculated. The derived layers include slope, curvature, Gray-Level Co-Occurrence Matrix (GLCM) features, Canny edge detection results.

To ensure robust classification, the data was split into tiles, with some used for training and others for validation to minimize spatial autocorrelation. The model was trained using AutoGluon, focusing on CatBoost and Neural Networks. The binary classification achieved a ROC-AUC value of 0.87 and a Matthews Correlation Coefficient (MCC) of 0.44. Subsequently, detected terraces were classified into three morpho-functional classes. Broad agricultural terraces were identified with high accuracy (0.84). Narrow agricultural terraces on steeper slopes with stone walls showed moderate performance (accuracy = 0.72). However, distinguishing narrow terraces built for reforestation from agricultural terraces in similar conditions proved challenging (accuracy = 0.46).

Our results demonstrate the potential for developing detailed, large-scale terrace datasets. This, in turn, opens promising perspectives to better assess soil erosion and other geohydrological processes at such scale.