Hugo Carlos-Martínez


Sessions

11-12
12:50
15min
A Bayesian SVC–SPDE approach to above-ground biomass estimation in tropical forests: response transformation and covariate dominance in Quintana Roo, Mexico
Katia Michelle Villarnobo Gonzalez, Hugo Carlos-Martínez, Jorge Omar López Martínez

1. Introduction

Reliable estimates of above-ground biomass density (AGBD) are essential for forest carbon accounting, REDD+, and conservation policies. Although Mexico's National Forest Inventory (INFyS) provides the most reliable field-based AGBD data, its limited spatial coverage and high acquisition costs constrain large-scale monitoring, particularly in tropical forests. Earth Observation (EO) products, such as ESA CCI Biomass and GEDI-derived canopy height, offer continuous spatial coverage but are affected by non-stationary relationships with AGBD and sensor saturation. Consequently, integrating EO and field data requires models that explicitly account for spatial dependence and uncertainty, making spatially varying coefficient (SVC) geostatistical models a suitable approach.

2. Research Gap
Models fitted over large, heterogeneous domains often borrow strength across broad ecological gradients and tend to shrink local predictions toward the national mean. While statistically efficient at the national scale, this can lead to regional miscalibration, particularly where Earth Observation (EO) covariates lose sensitivity. Both ESA Climate Change Initiative Biomass (CCI Biomass), which becomes weakly informative beyond roughly 150–200 Mg ha$^{-1}$, and GEDI-derived canopy height, which saturates near the upper end of tall tropical canopies, degrade precisely in dense tropical forests. Quintana Roo, with semi-evergreen tropical forest, secondary vegetation, coastal gradients, karstic substrates, and heterogeneous land-use histories, exemplifies this condition. A state-scale application therefore requires an independent modeling framework that jointly addresses response scale, data quality, and local covariate behavior, explicitly identifying where each EO source drives predictions.
3. Methodology

We develop such a reformulation for Quintana Roo within the SVC–SPDE framework. On a transformed response, the model is written in the standard additive form

$$ g{Y(\mathbf{s})}=[\alpha+\tilde{\alpha}(\mathbf{s})]+[\beta+\tilde{\beta}(\mathbf{s})]x_1(\mathbf{s})+[\eta+\tilde{\eta}(\mathbf{s})]x_2(\mathbf{s})+\epsilon(\mathbf{s}), $$

where $\tilde{\alpha},\tilde{\beta},\tilde{\eta}$ are mean-zero Matérn Gaussian fields approximated by the SPDE over a triangular mesh. The database combines INFyS plot AGBD, CCI Biomass, and GLAD-GEDI canopy height, harmonized around 2019, with covariates extracted as footprint-weighted means per plot. Because spatial covariance is inferred from the data, a single high-leverage record can distort local coefficient surfaces, so we applied a quality-control protocol—coordinate and unit verification, removal of duplicates and implausible values, and restriction to mainland forest. The SPDE mesh was confined to Quintana Roo rather than a national grid, improving computational efficiency.

The choice of $g$ matters once predictions return to the Mg ha$^{-1}$ scale, motivating our departure from the square-root response used nationally. Writing $\mu(\mathbf{s})$ for the linear predictor, the square-root link inverts by squaring, whereas the logarithmic link inverts by exponentiation:

$$ g=\sqrt{\cdot}: Y=\mu^2=\dots+2,\beta(\mathbf{s})\eta(\mathbf{s}),x_1x_2+\dots, \qquad g=\log: Y=e^{\alpha(\mathbf{s})},e^{\beta(\mathbf{s})x_1},e^{\eta(\mathbf{s})x_2}. $$

Squaring introduces covariate interaction terms, so the marginal effect of each covariate depends on the other and on the intercept, making EO effects entangled. The logarithm instead factorizes biomass into independent multiplicative terms, matching its power-law structure and preserving per-covariate effects. Zero-valued records were dropped, as $\log 0$ is undefined; since $\exp{E[\log Y]}\neq E[Y]$, predictions were back-transformed from posterior predictive samples. Accuracy was assessed with $R^2$, RMSE, MAE, and bias, calibration with leave-one-out Probability Integral Transform, and prediction structure through covariate-contribution surfaces $C_{\mathrm{CCI}}(\mathbf{s})=\beta(\mathbf{s})x_1(\mathbf{s})$ and $C_{\mathrm{GEDI}}(\mathbf{s})=\eta(\mathbf{s})x_2(\mathbf{s})$, with relative dominance $D_{\mathrm{CCI}}=|C_{\mathrm{CCI}}|/(|C_{\mathrm{CCI}}|+|C_{\mathrm{GEDI}}|)$.

4. Findings

The logarithmic formulation substantially outperformed the square-root one, returning an $R^2$ of 0.68 against 0.38, RMSE of 26.7 against 46.9 Mg ha$^{-1}$, MAE of 21.9 against 33.2 Mg ha$^{-1}$, and bias of 4.2 against 16.8 Mg ha$^{-1}$—a 78% gain in explained variance, a 43% reduction in RMSE, and a near four-fold reduction in relative bias, from 27.6% to 7.0%. These gains reflect the positive, right-skewed, heteroscedastic nature of biomass: aligning the linear predictor with multiplicative mechanisms prevents spatially varying coefficients from absorbing distortions arising from an unsuitable response scale. Applied locally, the PIT diagnostic confirms better-calibrated behaviour once variance is stabilised and anomalous records are removed.

Because the log scale keeps covariate effects separable, the contribution surfaces admit a clean reading that the squared scale does not. A coefficient map alone is misleading—a steep local slope contributes little where its covariate is near zero—so weighting each slope by the value its covariate takes reveals which EO source governs prediction at a given point. Mapping dominance over Quintana Roo exposes a spatial division of labor: zones where CCI Biomass carries the signal, where canopy height does, and where neither does and the latent field absorbs the residual baseline. This matters because similar canopy heights can correspond to different biomass densities depending on composition and degradation. Hunka et al. (2024) map spatially varying coefficients but stop there; carrying the analysis through to contribution and dominance turns a surface of parameters into an operational map of per-sensor reliability—and is coherent only under a response scale that does not entangle the covariates

5. Contributions

This study proposes an independent log-scale Spatially Varying Coefficient–Stochastic Partial Differential Equation (SVC–SPDE) formulation for tropical biomass that integrates quality control, a state-restricted mesh, and posterior-sample back-transformation into a unified workflow, improving performance in high-biomass regimes where Earth Observation (EO) covariates saturate. Its main contribution is the spatial decomposition of covariate effects using a logarithmic response that preserves separability on the natural scale, enabling identification of where each EO source governs prediction. The framework also supports posterior predictive aggregation over administrative and conservation units with explicit uncertainty, providing a transferable approach for biomass estimation.

Geographic Information Systems and Cartography
Multifuncional (Remoto)
11-13
11:15
15min
Dynamic forest fire hazard mapping with artificial intelligence: An integrated geospatial framework for the Trans‑Mexican Volcanic Belt
Adrian Israel Silva Cardoza, Daniela Moctezuma, Hugo Carlos-Martínez

Background
Forest fires are a central ecological process in many regions of Mexico, particularly in temperate and montane ecosystems where fire has historically shaped vegetation structure and regeneration dynamics. However, under current conditions of climate change, land‑use transformation, and expanding human pressure, fire occurrence increasingly leads to high‑risk scenarios for communities, ecosystems, and Protected Natural Areas (ANP). In the central region of the country, decades of fire‑exclusion policies have contributed to fuel accumulation, increasing the likelihood of severe events. At the same time, traditional fire‑use practices, regulatory constraints, and heterogeneous patterns of land occupation create spatially differentiated ignition conditions. Existing monitoring systems often rely on static or coarse‑resolution approaches that do not adequately represent the spatial variability of fuels, microclimates, or anthropogenic drivers. These limitations underscore the need for methodologies that integrate meteorological, fuels, anthropogenic, and topographic predictors within a machine‑learning framework capable of capturing the dynamic conditions that influence fire occurrence probability. Combining these components enables the construction of hazard maps that anticipate critical scenarios and support adaptive fire‑management strategies. This work was carried out with the support of the Mexican Fund for the Conservation of Nature (FMCN) and the Development Bank KfW, whose collaboration has strengthened ongoing efforts to advance geospatial tools for fire‑risk assessment and sustainable landscape management.
Objective
Develop and implement a methodology for generating forest fire hazard maps in the Central Region and Trans‑Mexican Volcanic Belt. The aim is to explain spatial patterns of fire hazard by evaluating methodological improvements, including the integration of multivariate machine‑learning algorithms, optimized predictor selection, and threshold calibration for operational decision‑making.
Materials and methods
The study area covers 164,203 km² across 13 states and 771 municipalities. It includes 19 federal ANP that represent key conservation areas. The region exhibits strong altitudinal gradients, diverse vegetation types, and marked climatic variability, all of which influence fuel availability and ignition patterns. Fire occurrence records from 2012–2024 were obtained from the national forest agency. A binary presence–absence dataset was constructed by filtering events within forested areas and generating pseudo‑absences outside a 1‑km exclusion buffer around each ignition point. This approach reduces spatial autocorrelation and ensures that absence samples represent distinct environmental conditions. The final dataset includes annual ignition points and corresponding absence samples, each associated with vegetation type and year of occurrence. The independent variables were organized into four thematic components. The anthropogenic component included distance to protected areas, settlements, and agricultural zones, as well as population density derived, representing human activities that drive most ignitions. The forest‑fuel component incorporated spectral indices from VIIRS imagery—such as GOSAVI, NBR, and NDVI—processed into 60‑day composites to reduce atmospheric noise and capture short‑term vegetation conditions. The topographic component included slope, terrain roughness, and related metrics derived from the national elevation model and resampled to 1 km to ensure spatial consistency. The meteorological component used Daymet reanalysis to compute 60‑day composites of vapor pressure deficit, accumulated precipitation, accumulated radiation, and temperature extremes, representing atmospheric conditions that influence fuel moisture and ignition likelihood.
The dataset was split into training (60%), validation (20%), and external testing (20%) subsets using stratified sampling to preserve class proportions. Variables were normalized to the [0,1] range when required by the algorithms. A multi‑step procedure was applied: (i) Univariate AUC analysis to identify predictors with the highest discriminative capacity, (ii) Collinearity assessment using Pearson correlation and variance inflation factors, (iii) Stability verification using Y‑Scrambling to ensure that predictive relationships were not random, (iv) Four predictors were selected for the final model: distance to ANP (adan), terrain roughness (teri), 60‑day vapor pressure water (vp), and spectral index gosavi.
Five algorithms were evaluated: Random Forest, XGBoost, Support Vector Machines with radial kernel, k‑Nearest Neighbors, and logistic regression. Models were trained using stratified 10‑fold cross‑validation. Performance metrics included AUC, accuracy, sensitivity, specificity, Matthews correlation coefficient, and F1‑score. The two best‑performing models (Random Forest and XGBoost) were optimized using grid search. Final evaluation was conducted on the external test set, including threshold optimization using the Youden index. The final model was applied to a multiband raster stack of predictors aligned in resolution and extent. Continuous probabilities were classified into five hazard levels. Spatial residuals were computed to identify areas of over‑ or under‑estimation.
Results and discussion
Random Forest achieved the highest performance, with an AUC of 0.906, followed by XGBoost with 0.895. Both models showed consistent behavior across folds, indicating stable generalization. Threshold optimization increased sensitivity from 0.720 to 0.807 for Random Forest, improving detection of ignition events with a moderate reduction in specificity. The selected predictors captured key spatial gradients: ignitions were more frequent within 50 km of ANP, in areas of intermediate terrain roughness, under low atmospheric moisture (vapor pressure near 500 Pa), and in vegetation with intermediate values, reflecting moderate biomass and canopy openness. The results highlight the interaction between climatic drivers, fuel conditions, and human activities. The inclusion of distance to ANP captured spatial gradients associated with land‑use restrictions and community fire‑use practices. The 60‑day meteorological composites effectively represented short‑term atmospheric conditions relevant to ignition. Limitations include positional uncertainty in fire records, coarse spatial resolution of Daymet, and the absence of high‑resolution wind and humidity data. Future work should integrate hot-spots, thermal anomalies, and alternative climate datasets such as CHIRPS or ERA5‑Land, as well as medium‑resolution spectral data from Sentinel‑2 or Landsat.
Conclusion
The study demonstrates the feasibility of integrating machine learning with meteorological, spectral, and topographic predictors to generate dynamic fire hazard maps at regional scale. The methodology is scalable and suitable for operational implementation. Future efforts should evaluate multiple temporal windows, incorporate satellite‑derived ignition proxies, and expand validation across additional regions.

Artificial Intelligence and Geographical Information Science
Video Room