Assessing fragmentation of tropical dry forest in northern colombia using ensemble machine learning

Tropical dry forests (TDF) are the most threatened type of tropical forest, with less than 10% of their original extent remaining globally. In Colombia, only about 8% of the original TDF persists, and in some regions, current estimates remain imprecise or incomplete. To address this gap, we applied a supervised classification framework based on a multi-model ensemble of machine learning algorithms to map TDF cover in a highly fragmented landscape in northern Colombia. We used four algorithms – Random Forest, Support Vector Machines (with polynomial and radial kernels), and Neural Networks – and combined their outputs using a majority voting strategy. This approach produced a high-resolution (10 m) map of TDF distribution and fragmentation patterns. The ensemble classifier outperformed individual models, achieving an F1 score of 0.82. We estimate that TDF currently covers 5.39% (64,645 ha) of its potential distribution in the study area, distributed across 11,742 patches—98.4% of which are smaller than 40 ha and 93% located within 50 meters of a disturbed edge. Our classification results diverge substantially from previous national and global forest maps. These findings highlight the urgency of conservation efforts and demonstrate the utility of high-resolution, ensemble-based remote sensing tools for improving forest monitoring and informing regional land-use planning.