Modelado predictivo de la incidencia del deterioro del arbolado urbano con aprendizaje automático y GIS: implementación de servicios geográficos web para una zonificación de manejo para la ciudad de Cali - Colombia

Abstract

The city of Santiago de Cali, the third most populated city in Colombia with more than 1.5 million inhabitants, faces a critical challenge in the management of its urban tree population. Of the 291,088 inventoried trees, more than 80% of the most important native species exhibit some form of deterioration, while annual maintenance costs exceed one billion Colombian pesos without achieving intervention on even 7% of the total population. In response to this issue, this study developed a predictive model based on data mining techniques (decision trees with k=10 cross-validation and class balancing) and Geographic Information Systems (GIS), integrating 22 deterioration-related variables (exposed roots, bacteriosis, girdling, improper pruning practices, among others) together with physical and environmental variables such as temperature, precipitation, humidity, air pollutants (PM10, PM2.5, NO2, O3, SO2), urban infrastructure, and socioeconomic conditions. The results demonstrated robust model performance, achieving F1-scores above 0.91 for species such as Pithecellobium dulce and Melicoccus bijugatus. The analysis identified annual rainfall, humidity, temperature, and atmospheric pollutants (particularly PM2.5 and NO2) as the most influential variables associated with tree deterioration. A probability zoning map of deterioration risk was generated using a 50 × 50 m grid covering the 12,019.6 ha urban area, revealing that more than 60% of the territory presents a high probability of deterioration for most species. Nevertheless, specific areas located mainly in southern and central Cali, associated with parks and green areas, exhibited low deterioration probability for species such as Roystonea oleracea, with approximately 4,741 ha classified as low-risk zones. Finally, the study published an interactive web-based geovisor through ArcGIS Online Geovisor, enabling species-specific consultation of management zoning maps and unsuitable planting areas. This platform constitutes a decision-support tool for environmental authorities such as DAGMA and CVC. The study concludes that the integration of data mining and GIS provides a replicable methodology for evidence-based urban forestry management and recommends improving pruning practices, evaluating more advanced machine learning models (e.g., Random Forest and XGBoost), and implementing permanent interinstitutional strategies for the preservation of urban trees.