New Models Map Wildfire Risk in Iran’s Protected Forests

Researchers have developed high-resolution maps detailing burn probability and wildfire risk across six protected areas in Golestan Province, northeastern Iran. Published in Geoenvironmental Disasters, the study integrates fire spread simulations with machine learning to identify zones most susceptible to wildfires within the Hyrcanian temperate forest region. The work aims to provide operational tools for conservation authorities and fire managers to prioritize mitigation efforts in these ecologically sensitive landscapes, which have faced increasing wildfire activity in recent decades.

The study focused on six distinct protected sites: Golestan National Park (GNP), Loveh, Upper-Zav, Lower-Zav, Hyrcania, and Jahannama. These areas represent a mix of sub-humid Hyrcanian forests and semi-arid Irano-Turanian steppes, creating a complex mosaic of vegetation and fuel loads. GNP, the first national park designated in Iran, spans three provinces and serves as a critical habitat for biodiversity. The other sites, including the smaller forest-protected regions of Loveh and the Zav areas, are attached to or adjacent to GNP, while Jahannama and Hyrcania are larger protected zones in the southern mountains of the province.

To estimate spatial burn probabilities, the team utilized the Minimum Travel Time (MTT) fire growth algorithm within the FlamMap simulation platform. This deterministic approach simulates how fires spread across a landscape by calculating the time it takes for fire to travel from ignition points to every other cell. The researchers used 1,000 random ignition points distributed according to historical ignition density patterns. Each simulation ran for a fixed 12-hour duration, representing the typical length of large fires in the region, and incorporated historical mean weather conditions, including prevailing east and northeast winds that account for over 60% of recorded wind directions.

Custom fuel models were derived from field measurements of dead and live fuels, including 1-hour, 10-hour, and 100-hour dead fuels, as well as live herbaceous and woody components. These models were based on previous work in the study area and were used to represent the specific vegetation structures found in the protected zones. The simulation outputs provided a conditional burn probability for each 30-meter grid cell, ranging from zero, indicating the cell never burned in any simulation, to one, indicating it burned in every simulation. This metric reflects the likelihood of a cell burning given that a fire has started somewhere in the landscape.

For wildfire risk assessment, the researchers employed Support Vector Regression (SVR), a machine learning algorithm known for its ability to handle nonlinear relationships in high-dimensional data. The model used the Radial Basis Function (RBF) kernel, which is effective for environmental modeling tasks involving complex interactions between variables. The SVR model was trained on 70% of the historical fire ignition data from 2013 to 2024 and tested on the remaining 30%. Key hyperparameters, including gamma, capacity, and epsilon, were optimized using a grid search algorithm with 10-fold cross-validation to ensure the model’s predictive accuracy and generalization capability.

The results revealed significant spatial variations in burn probability across the six sites. In Golestan National Park, the highest burn probability was recorded at elevations below 500 meters, reaching a maximum of 0.096 in specific zones. This suggests that lowland areas are particularly vulnerable, likely due to denser fuel continuity, drier microclimates, or greater exposure to ignition sources. Conversely, burn probabilities declined with increasing elevation, dropping to near zero in zones above 1,500 meters. Similar trends were observed in Loveh, where lower elevations showed moderate risk, while higher elevations remained largely unaffected by simulated fires.

Fuel type also played a critical role in determining burn probability. Areas with fine and continuous fuels, represented by fuel models FM1, FM3, and FM4, exhibited the highest burn probabilities. These fuel types, common in conifer plantations and deciduous forests with significant undergrowth, support rapid ignition and fire spread. In Hyrcania, for instance, FM1 recorded the highest burn probability among all fuel models, indicating that fuel beds composed of fine herbaceous or shrub materials are particularly susceptible to fire. In contrast, fuel models associated with heavy dead woody material or sparse vegetation, such as FM10, showed near-zero burn probabilities in most areas, suggesting these zones are less likely to sustain active fire spread under normal conditions.

The SVR-based fire risk maps categorized the study areas into four risk classes: low, moderate, high, and very high. The models achieved high overall classification accuracies, ranging from 88% to 95% across the different regions. In areas classified as high or very-high risk, 54% to 72% of observed historical fire ignitions were located, demonstrating the model’s effectiveness in identifying zones of concern. The highest predictive power was observed in Hyrcania, where the overall accuracy reached 95%. This performance indicates that the SVR framework, when calibrated with region-specific parameters, can reliably predict wildfire susceptibility in complex forested landscapes.

Topographic factors, including elevation, aspect, and slope, significantly influenced the spatial patterns of wildfire risk. In most protected areas, southern and eastern aspects recorded higher burn probabilities, likely due to increased solar radiation and reduced fuel moisture. Steeper slopes also showed higher burn probabilities, as slope gradient enhances fire intensity and spread rate by preheating upslope fuels. However, exceptions were noted in Jahannama and Hyrcania, where burn probabilities increased with elevation, suggesting that specific fuel structures or dry, exposed ridges in these mountainous areas maintain elevated fire potential despite higher altitudes.

The study acknowledges several limitations, including the assumption of constant fuel moisture and wind conditions within each simulation scenario, which may not fully reflect real-world variability. The resolution of the datasets could also limit the ability to capture fine-scale fuel variability, and spatial autocorrelation among data points may affect model accuracy. Future research should incorporate dynamic weather scenarios and higher-resolution fuel datasets to enhance the robustness of these models. Additionally, integrating real-time meteorological data and climate change projections could lead to the development of dynamic fire risk forecasting systems.

This research provides a localized wildfire risk assessment framework for the Hyrcanian forests, addressing a gap in existing literature that often relies on static fire inventories or global models. By combining fire behavior modeling with data-driven approaches, the study offers a transferable methodology for predicting wildfire risk in other ecologically sensitive regions. The resulting maps and risk classifications can support operational fire risk zoning, helping resource managers allocate suppression resources more effectively and improve long-term conservation planning in the face of changing climate and land use conditions.

A detailed examination of vegetation fuel complexes revealed how specific forest stand characteristics govern wildfire likelihood across the protected territories. Elevated burn probabilities were strongly concentrated in stands dominated by fine surface fuels, including fallen needles, dried litter, and slender twigs. These conditions frequently occurred in conifer plantations and deciduous canopies harboring dense understory vegetation, particularly in tracts populated by Mediterranean cypress (Cupressus sempervirens), Turkish pine (Pinus brutia), chestnut-leaved oak (Quercus castaneifolia), and Persian ironwood (Parrotia persica). In these stands, abundant one-hour and ten-hour dead fuels promote rapid ignition and continuous surface fire propagation. Conversely, fuel models representing compact, moisture-retaining fuels or sparse ground cover exhibited minimal burn likelihood. In fuel model FM10, burn probabilities dropped to 0.000 in Upper-Zav, 0.002 in Golestan National Park, and 0.005 in Jahannama. However, transitional fuel beds captured by fuel model FM11 generated notable fire potential, recording an average probability of 0.025 in Jahannama and 0.014 across the Zav reserves.

Topographic analyses highlighted pronounced contrasts in burn probability across distinct elevation belts and slope inclinations within individual reserves. In Golestan National Park, the mean burn probability reached 0.032 at elevations below 500 meters, decreased to 0.017 between 500 and 1,000 meters, and approached zero above 1,500 meters. Conversely, Upper-Zav and Lower-Zav displayed an isolated mid-elevation peak of 0.017 between 1,000 and 1,500 meters, whereas Hyrcania recorded its peak probability of 0.026 within the 1,500 to 2,000-meter zone. Aspect also produced marked local variations. On the sun-exposed southern and southeastern slopes of Upper-Zav and Lower-Zav, probabilities reached 0.015 and 0.013, respectively, whereas Hyrcania exhibited its highest burn probabilities on east-facing slopes at 0.029 and northeast-facing slopes at 0.024. Across all reserves, steep terrain accelerated fire behavior; in Golestan National Park, burn probability increased from 0.002 on slopes under 10 percent to 0.009 on slopes exceeding 60 percent, while Hyrcania reached 0.022 on the steepest inclines.

The modeling framework also accounted for spatial predictors and algorithmic optimization across three designated regional sectors: Golestan National Park combined with Loveh and the two Zav reserves, Hyrcania, and Jahannama. Environmental layers, including distances to roads, agricultural settlements, and recreational campsites, were integrated at a basic modeling unit of one hectare. In configuring the Support Vector Regression models, optimal hyperparameter values were established through ten-fold cross-validation repeated over 1,000 iterations. The parameter search bound capacity values between 1 and 20 while varying epsilon from 0.1 to 1, with gamma defined as the inverse of the total number of predictors. Beyond closed forest canopies, post-harvest cropland interfaces and agricultural residues exhibited heightened seasonal flammability, underscoring how human land-use borders interact with natural woodland fuels to elevate fire susceptibility during dry harvesting periods.

Subject of Research: Burn probability and wildfire risk modeling of the protected areas in Golestan Province, NE Iran

Article Title: Burn probability and wildfire risk modeling of the protected areas in Golestan Province, NE Iran

Article References: Jouibary, S. S., Jahdi, R., Khalaf, W. A., & Eshaghi, M. A. (2026). Burn probability and wildfire risk modeling of the protected areas in Golestan Province, NE Iran. Geoenvironmental Disasters, 13(1), Article 32. https://doi.org/10.1186/s40677-026-00379-3

Image Credits: AI Generated

DOI: 10.1186/s40677-026-00379-3

Keywords: Burn, probability, wildfire, risk, modeling, protected, areas, Golestan, Province, Iran, scientific research