AI Maps Where Fire and Habitat Loss Threaten the World’s Largest Lizard

The Komodo dragon, the largest lizard alive today, survives on a handful of small islands in eastern Indonesia, and its future is narrowing with every passing decade. In 2021 the International Union for Conservation of Nature upgraded the species from Vulnerable to Endangered after finding that fewer than 1,400 adults remain, no single subpopulation holds more than 500 individuals, and the entire species occupies just 809 square kilometres of severely fragmented habitat. A new study published in Environmental and Sustainability Indicators now offers the most detailed machine-learning map yet of where those dragons live well and where they are most likely to burn, combining satellite remote sensing, open biodiversity data and ensemble artificial intelligence to guide conservation decisions inside Komodo National Park.

The research team, led by Anjar Dimara Sakti of Institut Teknologi Bandung, built a habitat suitability index that treats the dragon’s ecological needs as a family of linked indicators rather than a single statistical guess. Because Komodo dragons are age-structured predators, the model separately characterised adult and juvenile requirements. Adults hunt ungulates such as Timor deer, water buffalo and feral pigs across open savanna, while juveniles are arboreal, feeding on small lizards, large insects and snakes in taller trees where they can escape cannibalism by their own kind. The framework therefore modelled three indicators in parallel: the distribution of adult prey, the distribution of juvenile prey, and the ecological conditions where the lizards themselves have been recorded.

To train the habitat model, the researchers drew 144 verified occurrence records of Varanus komodoensis from the Global Biodiversity Information Facility, filtering the citizen-science data by species identity, geography and coordinate uncertainty below one kilometre, a threshold matched to the animal’s daily roaming distance. Because absence data do not exist for such a wide-ranging predator, the team generated pseudoabsence points using a multi-criteria analysis that scored the landscape for adult prey, juvenile prey and ecological suitability, retaining only the lowest-scoring locations as clearly unsuitable ground. Those 169 pseudoabsences, balanced against the presences, fed into a Maximum Entropy model alongside seventeen environmental predictors spanning topography, climate, land cover, tree cover, canopy height and satellite-derived surface temperature.

The results paint a precise picture of dragon geography. Roughly 36 percent of Komodo National Park, about 20,885 hectares, falls into the high or very high suitability classes, with the best habitat concentrated on the northern side of Rinca Island, across much of Padar, and in low-lying, flat savanna with moderate tree cover. The model achieved an area under the receiver operating characteristic curve of 0.937, outperforming earlier Komodo habitat studies that reached 0.776 and 0.842 with narrower predictor sets. Unsurprisingly for a ground-dwelling ectotherm, surface temperature and the transitional mosaic between open grassland and closed forest emerged as decisive variables: the lizards shuttle between sun and shade to regulate body temperature, and the same mosaic sustains the deer and pigs that anchor the food web.

Habitat alone, however, does not determine survival, and the study’s most consequential move was to treat wildfire susceptibility as an inseparable part of the conservation equation. Komodo National Park’s lowland savanna, dominated by drought-adapted grasses and scattered lontar palms, cures into highly flammable fuel during the long May-to-October dry season, and fires burned 170 hectares of savanna between 2013 and 2021. The team trained three tree-based machine-learning algorithms, classification and regression trees, gradient tree boosting and random forest, on 150 wildfire points extracted from Sentinel Hub together with 250 non-fire points, using predictors that included elevation, slope, wind speed, land surface temperature, Sentinel-2 vegetation and moisture indices, and Sentinel-1 radar backscatter.

Random forest proved the strongest individual learner, with a mean AUC of 0.885 across 100 repeated data splits, followed by gradient tree boosting at 0.863. The researchers were candid about the ensemble’s limits: averaging all three algorithms equally dragged performance down because the weakest learner contributed as much as the strongest, and a two-model ensemble of random forest and gradient boosting scored higher at 0.878. More importantly, the team tested how well the models transferred beyond their training geography. Under spatially blocked cross-validation and leave-one-fire-out validation, in which each of the three mapped burn polygons was withheld in turn, discrimination fell from 0.912 on random partitions to 0.799 and 0.735 respectively, a sobering reminder that apparent accuracy can inflate when fire records cluster in accessible terrain.

The wildfire model’s explanatory power came from interpretable machine learning. Permutation importance and SHAP analysis converged on a consistent story: susceptibility rises as vegetation greenness (NDVI) and radar backscatter decline, as the burned-area index NBR+ increases, and as wind speed and maximum land surface temperature climb. In plain terms, the flammable landscape is sparse, dry, low-biomass savanna under warm, windy conditions, precisely the coastal lowlands the dragons favour. Slope and elevation, by contrast, contributed little once the vegetation indices were included, suggesting that topography shapes fire risk mainly through the vegetation it supports. The model also exposed a gap between official accounts and satellite reality: three fires officially reported at 10 hectares each in 2018 and 2021 measured 102.93, 184.63 and 300.07 hectares in Sentinel-2 imagery, with one blaze suspected of being set by poachers to divert rangers.

Overlaying the two surfaces produced the study’s central deliverable, a bivariate conservation priority map. About 58 percent of the park sits in the lowest wildfire vulnerability class, but 11 percent, some 7,200 hectares, is highly vulnerable, concentrated along low-elevation coastal savanna. Critically, 2,800 hectares, five percent of the park, combine high habitat suitability with high fire risk, and another 3,500 hectares of moderate suitability face the same threat. When the researchers overlaid seven Komodo movement paths from historical field monitoring, most dragon ranges fell in low-vulnerability zones, but the paths of two individuals crossed medium-to-high fire exposure, and one path contained the single largest share of high-suitability, high-risk ground. The analysis also revealed a mismatch between ecology and management: the park’s core protection zone holds less high-quality, low-risk habitat than the adjacent jungle zone, prompting the authors to recommend extending core-zone boundaries into those safer, high-suitability patches.

The study is explicit about its own uncertainties. The wildfire samples derive from only three burn events rather than 150 independent fires, the canopy height layer dates from 2005, and adding human-accessibility predictors produced an apparent accuracy gain that collapsed under spatial validation, indicating that ignition sources and flammable terrain coincide too closely in this small park to separate statistically. The authors stress that the priority maps should be treated as decision-support hypotheses requiring field confirmation, not precise boundaries for allocating suppression resources. Yet the framework’s real significance may lie beyond Komodo. By coupling an ecologically informed, multi-indicator species model with a threat-susceptibility model and a formal zoning evaluation, all built from free satellite and biodiversity data, the approach offers a transferable template for protecting other range-restricted species in seasonally dry, fire-prone landscapes, from island reptiles to fragmented mammal populations across Southeast Asia.

For the dragons themselves, the practical prescriptions are concrete: early dry-season fuel management and firebreaks around movement corridors in the coastal savanna overlap, tighter ignition controls given the suspected human role in past fires, seasonal restrictions on tourism in the highest-risk zones, and periodic re-running of both model layers so that zoning evolves with new data. With climate change degrading vegetation quality, prey populations declining on multiple islands, and plans to push tourism on Rinca toward more than a million visitors a year, the world’s largest lizard now has something it has never had before: a data-driven map of exactly where its most valuable ground and its most dangerous threat intersect.

Subject of Research: Machine learning-based habitat suitability and wildfire susceptibility modelling for Komodo dragon conservation planning in Komodo National Park, Indonesia

Article Title: Machine learning-based Komodo dragon conservation planning using habitat suitability and wildfire risk

Article References: Sakti, A. D., Zakiar, M. R., Rosleine, D., Artaningh, F., Santoso, C., Pramudya, A. D., Candra, D. S., Wijaya, S., & Wikantika, K. (2026). Machine learning-based Komodo dragon conservation planning using habitat suitability and wildfire risk. Environmental and Sustainability Indicators, 32, Article 101540. https://doi.org/10.1016/j.indic.2026.101540

Image Credits: AI Generated

DOI: 10.1016/j.indic.2026.101540

Keywords: Komodo dragon, machine learning, habitat suitability, wildfire susceptibility, species distribution model, MaxEnt, random forest, remote sensing, Komodo National Park, conservation planning, Sentinel-2, biodiversity