Modeling forest restoration potential in the Scottish Highlands using multiple machine learning approaches

Abstract Natural range contraction and millennia of anthropogenic disturbance have led to a steady decline in stands of Caledonian Pine Forest (CPF) across the Scottish Highlands. Much of the land is now dominated by short-stature shrubs, with fragmented areas of native forest and commercial plantation. Several surmountable barriers exist to large-scale reforestation of the CPF, including tensions with existing economies, land ownership patterns, and the uncertainties posed by a changing climate. To address these challenges and support decision-making, we developed a data-driven approach to identify optimal sites for native forest restoration across the CPF. We trained, validated, and deployed five machine learning classification models – multilayer perceptron, naïve Bayes, random forest, support vector machine and XGBoost – to predict which sites across the CPF ecoregion were most suitable for one of three broad native forest community types: Scots pine, oak woodland and birch woodland. In the least restrictive reforestation scenario, we identified a total of 844,339 hectares of potential reforestation area while the most restrictive reforestation scenario identified 210,703, hectares. Birch, Scots pine and then Oak ranked most to least for predicted sites. Among the models, XGBoost demonstrated the highest predictive power using area under the receiver operating characteristic (AUC = 0.974, Accuracy = 0.918) while Naïve Bayes performed the least effectively (AUC = 0.794, Accuracy = 0.655). Our findings provide a spatially explicit foundation for prioritising reforestation efforts, enabling stakeholders to maximise ecological gains while navigating competing land use pressures.