Mapping agricultural terrace suitability in Mediterranean mountain environments using a positive–unlabelled classification framework

Abstract

Agricultural terraces in Mediterranean mountain environments provide a sustainable means of farming on steep terrain while delivering essential ecosystem services. However, land desertification driven by land abandonment and a changing climate raises new concerns for the sustainability of these environments. Land suitability analysis (LSA) provides a valuable tool to support decision-making for sustainable management and potential expansion of terrace agriculture. Traditional LSA approaches typically require fully labelled dataset, but in many real-world applications only a fraction of positive examples is available, with the rest unlabelled. This study aims to present an integrated predictive modelling framework that combines GIS with data-driven Machine Learning (ML) techniques, capable of learning from positive and unlabelled datasets for LSA. The proposed framework was applied to develop a terrace suitability map for Cyprus’ Troodos Mountains. A 5-m DEM was processed to extract the mountain area, with elevation ≥500m and slopes ≥15%, defining the study area. Crop plots registered under the Single Area Payment Scheme of the European Common Agricultural Policy were used to classify the study area into Terrace-Present (TP) and Terrace-Absent (TA) cells, with TP serving as labelled positive and TA as unlabelled samples. A two-step ML approach was applied, first identifying reliable negatives from TA cells, then using these with TP cells for suitability prediction. Despite a high class imbalanced between positive (3.4%) and unlabelled dataset (96.6%), the developed PU classifier achieved a Recall of 84.6%, Precision of 81.5%, and an F1 score of 83%, demonstrating robust and balanced performance. The resulting suitability map identified approximately 7,000 ha of land in the highest suitability class, indicating potential for future terrace development. Feature importance analysis identified land cover as the most influential parameter accounting for 23.9% of the total mean SHAP value, while terrain slope and tree cover density contributed 15.0% and 14.9%, respectively. Comparative analysis between 2017 and 2024 revealed abandonment of terraced agricultural land (29% decrease) as well as revitalization (12% increase). The resulting suitability map and accompanying data layers are accessible through a Google Earth Engine application, aiming to support informed decision-making for sustainable landscape planning.

This research has received financial support from the REACT4MED Project (GA 2122), which is funded by PRIMA, the Partnership for Research and Innovation in the Mediterranean Area, a Programme supported by Horizon 2020, the European Union’s Framework Programme for Research and Innovation.