Evaluación de modelos de aprendizaje automático para la determinación de zonas homogéneas físicas urbanas

Abstract

This research comparatively evaluates the performance of the machine learning models Random Forest, Support Vector Machine (SVM), and XGBoost in the delimitation of urban Physical Homogeneous Zones (PHZ) in the Bonanza neighborhood, Engativá locality, Bogotá D. C. The three algorithms were trained under identical conditions using 14 predictor variables derived from an RGB orthophoto (visible spectrum bands, spectral indices, and textural variables), employing as reference the delimitation obtained through the conventional methodology. Performance was assessed using metrics derived from the confusion matrix (Accuracy, Kappa coefficient, F1-Score, and Balanced Accuracy), internal homogeneity analysis, inter-model agreement, and spectral separability. The results showed a moderate and similar performance among the algorithms, with none satisfactorily reproducing the reference zoning, as only two to three of the eleven PHZ were correctly identified. The study concludes that the main limitation lies not in the algorithm employed but in the spectral information itself, given that PHZ integrate physical, urban-planning, and regulatory criteria that are not observable through remote sensors. Thus, remote sensing and machine learning constitute support tools to complement cadastral processes, but they do not replace the conventional methodology for determining PHZ.