From Ecological to Automated: Classification of High-Andean Vegetation, the case of "Altos de Potosí" - Colombia

Vegetation structure can be classified using remote sensing imagery, either through visual interpretation grounded in landscape ecology approaches or via automated digital analysis models. As these techniques continue to advance, it becomes essential to evaluate model performance using reliable data and detailed ecological maps. In this study, the Altos de Potosí region, located in Guasca, Colombia, at 3000 meters above sea level in the high Andean zone, was ecologically characterized based on field data. This enabled the construction of a detailed vegetation map, with units manually delineated on Google Satellite images. Seven natural vegetation classes and two classes corresponding to transformed systems were identified. These data were then used to train two machine learning models (Random Forest and K-Neighbors) to classify vegetation based on multispectral imagery from Landsat-8 and Sentinel-2 satellites. In the analysis using Landsat-8 imagery, the K-Neighbors model showed the best performance, with an overall accuracy of 58.75% and an F1-Score of 58.21%. In comparison, Random Forest achieved an overall accuracy of 51.16% and an F1-Score of 50.75%. With Sentinel-2 imagery, Random Forest outperformed K-Neighbors, attaining an accuracy of 70.00% and an F1-Score of 69.68%, while K-Neighbors yielded an overall accuracy of 69.22% and an F1-Score of 69.16%. Model performance was particularly low in small landscape units located in areas with high ecosystem diversity. Nevertheless, all vegetation classes were represented in both the ecological reference map and the maps generated by the models.