Edgar Libardo Barahona Cantor

Ingeniero Catastral y Geodesta con experiencia en análisis de datos espaciales, imágenes den satélite de baja, media y alta resolución, modelación estadística y gestión de información geográfica y alfanumérica. He trabajado en el desarrollo de soluciones orientadas al monitoreo ambiental, análisis multitemporal, evaluación de riesgos y análisis territorial, incluyendo procesos asociados al catastro y la valoración predial. Con capacidad para gestionar y dar seguimiento a proyectos mediante herramientas colaborativas de planificación. Actualmente curso estudios de maestría en Geomática, complementando mi formación en análisis espacial, programación y manejo de bases de datos para transformar datos complejos en información útil parala toma de decisiones en distintos contextos.


Sesión

12/11
13:05
15minutos
Spatial Modeling of Deforestation Susceptibility Using INLA-SPDE: A Comparative Assessment Against Traditional Logistic Regression in Guaviare, Colombia
Edgar Libardo Barahona Cantor

Tropical deforestation represents one of the most critical environmental threats of the 21st century, with direct implications for biodiversity, the hydrological cycle, and regional and global carbon balance. In Latin America, Colombia stands as one of the countries facing the greatest pressure on forest cover, with the department of Guaviare ranking among the territories with the highest rates of forest loss in the country, driven by the expansion of the agricultural frontier, extensive cattle ranching, and the presence of illicit crops. These pressures have resulted in the loss of approximately 452,000 hectares of forest cover between 2001 and 2024, equivalent to 8.6% of the department's total forest area. In this context, spatial modeling of deforestation susceptibility emerges as a fundamental tool for guiding conservation strategies and land-use planning policies. However, most existing studies apply methods that assume spatial independence among observations, ignoring that deforestation operates as a contagion process with strong spatial autocorrelation — a limitation that leads to biased estimates of conditioning factors and susceptibility maps with reduced discriminative precision.

This study evaluated whether explicitly incorporating spatial autocorrelation through the Integrated Nested Laplace Approximation with a Stochastic Partial Differential Equation approximation (INLA-SPDE) improves the deforestation susceptibility mapping capacity compared to a traditional logistic regression (LR) model. The analysis was conducted on 113,504 deforestation and non-deforestation points in the department of Guaviare, Colombia, for the period 2012–2024, derived from the forest cover loss catalog of the Institute of Hydrology, Meteorology and Environmental Studies (IDEAM). Twelve explanatory variables of biophysical and anthropogenic nature were used: slope, elevation, aspect, mean annual precipitation, distance to roads, distance to agricultural areas, distance to rivers, distance to drainage systems, distance to populated centers, distance to protected areas, population density, and mean annual temperature.

The exploratory data analysis — both statistical and spatial — confirmed the relevance of the spatial modeling approach. The global Moran's I index yielded a value of I = 0.830 (p < 2.2 × 10⁻¹⁶), demonstrating significant positive spatial autocorrelation. LISA indicators identified well-defined High-High and Low-Low spatial clusters, while Ripley's L(d) and F(d) functions confirmed a consistent spatial clustering pattern across multiple scales, reinforcing the evidence of non-random spatial structure in the deforestation phenomenon. This body of evidence justified the use of models that explicitly incorporate spatial dependence. Correlation and multicollinearity analysis (VIF) determined the exclusion of mean temperature — the variable with the highest redundancy relative to precipitation (ρ = 0.93) — and motivated a log transformation of variables with highly skewed distributions, followed by Z-score standardization of all covariates prior to model fitting.
Data partitioning was performed differently for each model: a stratified random split (80/20) was applied for the LR model, while spatial block cross-validation using 10,000 × 10,000 m blocks was implemented for INLA-SPDE, consistent with the estimated spatial correlation range of approximately 5,000 m. This approach was adopted to prevent the overestimation of predictive performance associated with spatial contiguity between training and test sets, ensuring that model evaluation reflects genuine generalization capacity toward unobserved areas.

Results demonstrate the superiority of the spatial model across all evaluation criteria. The INLA-SPDE model achieved an AUC of 0.9705 compared to 0.9544 for the LR model, representing an improvement of 1.61 percentage points attributable exclusively to the incorporation of the latent spatial component. The estimated Gaussian field presented a correlation range of 5,290 m and a standard deviation of σ = 3.38, indicating a relevant residual spatial structure not fully captured by the observed covariates. Evaluation using the SCAI index revealed that the INLA-SPDE model concentrated 88.20% of the observed deforestation within the very high susceptibility class (SCAI = 0.205), compared to 73.20% for the LR model (SCAI = 0.241), with a sharper separation between classes at both extremes of the susceptibility distribution.

The comparative coefficient analysis identified three behavioral groups. A first group — distance to agricultural areas, distance to roads, mean precipitation, and distance to drainage systems — maintained or intensified its effect after controlling for spatial dependence, establishing these variables as structurally stable determinants of deforestation in Guaviare. A second group — distance to rivers and elevation — collapsed toward zero or reversed sign in INLA-SPDE, revealing that their effects in the logistic model were largely spurious, acting as proxies for unmodeled spatial heterogeneity rather than reflecting genuine causal relationships. A third group — slope and distance to protected areas — exhibited sign reversals accompanied by changes in statistical significance, uncovering real effects that had been masked by uncontrolled autocorrelation in the classical specification.

The overall performance of both models surpasses the range reported in the literature for classical logistic methods (AUC between 0.68 and 0.87), machine learning approaches such as Support Vector Machine (AUC = 0.907) and artificial neural networks (AUC between 0.869 and 0.921), and multicriteria approaches such as VIKOR (AUC = 0.938). The advantage of the INLA-SPDE model is particularly significant in the Colombian Amazonian context, where deforestation exhibits a pattern of contiguous spatial expansion from pre-existing deforestation nuclei — a structure that the latent Gaussian field efficiently captures through its estimated correlation range, and that classical logistic regression fundamentally cannot represent under its independence assumption.

These results position INLA-SPDE as a methodologically robust and predictively superior alternative for deforestation susceptibility mapping in territories with marked spatial structure. Furthermore, they open the possibility of extending this framework toward spatiotemporal models capable of projecting the future dynamics of the forest frontier in the department of Guaviare, contributing to the development of evidence-based early warning systems for forest conservation in the Colombian Amazon.

Inteligencia Artificial y Ciencias de Información Geográfica.
Sala Geomática 7