The research team, led by Sudeep Jogi Kanwar and Milan Dhakal of the Agriculture and Forestry University in Nepal together with Ashok Parajuli of the Division Forest Office, focused on Rasuwa because it embodies the mounting dangers facing mountain ecosystems worldwide. The district spans roughly 154,400 hectares, with elevations ranging from 614 to 7,227 meters, and about 68 percent of its territory lies above 3,000 meters. Long-term records from the Dhunche meteorological station show maximum temperatures rising by about 0.05 degrees Celsius per year between 1989 and 2018, while winter precipitation has declined. This elevation-dependent warming is eroding the natural flammability barriers that once kept fire out of the high mountains, and the region has already witnessed catastrophic blazes, including a 2012 wildfire that burned unchecked for months near the Gatlang forest.

To build the model, the researchers assembled eleven predictive layers spanning four categories of fire drivers. Biophysical factors included the 2022 National Land Cover Map of Nepal and the Normalized Difference Vegetation Index, or NDVI, derived from Landsat imagery, which captures vegetation density and moisture status. Topographic variables—elevation, slope, aspect, and the topographic wetness index—were calculated from the Shuttle Radar Topography Mission digital elevation model. Anthropogenic pressure was represented by Euclidean distance layers computed from settlement and road datasets, while climatic conditions came from WorldClim precipitation data, MODIS land surface temperature composites spanning 2000 to 2024, and long-term average wind speeds from the Global Wind Atlas. All layers were resampled to a common 30-meter resolution to reduce uncertainty arising from mismatched spatial scales.

The fire record itself came from the Visible Infrared Imaging Radiometer Suite aboard the Suomi National Polar-orbiting Partnership satellite, which detects active fires at 375-meter resolution—finer than the 1-kilometer MODIS product and therefore better suited to catching small, scattered mountain ignitions. The team extracted incidents between 2012 and 2024, retained only detections with confidence levels above 30 percent, and applied a spatial thinning procedure requiring a minimum 100-meter separation between points. That filtering left 757 well-distributed fire locations for model training, a critical step because clustered records can bias machine learning classifiers toward overrepresented landscapes.

Before training, the researchers screened their predictors for multicollinearity using variance inflation factors, a standard diagnostic that flags variables whose information overlaps so heavily that the model cannot attribute influence cleanly. Precipitation failed the test spectacularly, with a VIF of 20.14 against a threshold of 10, and was excluded. The authors attribute this to the orographic precipitation mechanism of the Hindu Kush Himalaya, in which monsoon moisture interception is controlled almost entirely by topographic position—meaning elevation already encodes much of the rainfall signal. The remaining ten variables passed, and the team proceeded to train a Random Forest classifier, an ensemble method introduced by Leo Breiman that grows hundreds of decision trees on randomly resampled subsets of the data and averages their votes.

Model tuning was systematic rather than ad hoc. Using five-fold cross-validation on 80 percent of the data, the team searched a grid of hyperparameters covering the number of trees, the minimum size of terminal nodes, and the number of variables considered at each split, selecting the combination that maximized the area under the receiver operating characteristic curve. The final configuration achieved an AUC of 0.90 on the held-out 20 percent test set, with a standard deviation of just 0.018 across folds. The True Skill Statistic, a metric favored in ecological modeling because it remains reliable under class imbalance, reached 0.67, and a Brier score of 0.12 confirmed that the model’s predicted probabilities were well calibrated rather than merely well ranked.

When the researchers interrogated which variables drove the predictions, elevation emerged as the single most important factor, followed by land surface temperature, NDVI, and proximity to settlements. Partial dependence plots revealed the shape of these relationships in striking detail. Fire probability peaked at lower elevations between roughly 1,000 and 2,500 meters, then plateaued with notable fluctuations between 2,500 and 4,000 meters before stabilizing near 4,200 meters—a pattern suggesting that subalpine forests and shrublands, long assumed to be too cold and wet to burn, now carry genuine risk. NDVI values between 0.4 and 0.6, corresponding to open needle-leaved forests rich in resinous, low-moisture conifer fuels, proved most dangerous. Fire likelihood also climbed steadily with temperature above 15 degrees Celsius and was highest on slopes of 30 to 60 degrees and on sun-exposed south-facing aspects, where fuel dries fastest in the Northern Hemisphere.

The resulting risk map, classified into five zones using the Natural Breaks method, proved operationally meaningful rather than statistically arbitrary: 87.47 percent of all recorded fire incidents fell within the very high risk class, while fewer than 1 percent landed in the low and very low categories combined. In total, 19,523 hectares of forest face high or very high risk. The geography of danger is uneven. Kalika Rural Municipality has the largest proportion of high-risk forest at 53.7 percent, while Gosaikunda has the least at 19.5 percent, though Gosaikunda holds the greatest absolute area of very high risk at 6,235 hectares. The western flank of Langtang National Park stands out as a hotspot, combining low elevations, high land surface temperatures, and rugged terrain that hampers suppression once ignition occurs. Temporally, the picture is equally concentrated: 91.86 percent of incidents occurred between January and April, April alone recorded 283 fires, and the jump from 24 incidents in 2022 to 223 in 2023 signals an accelerating trend. Notably, about 30 percent of fires occurred above 3,000 meters, with detections as high as 4,525 meters.

The implications reach beyond fire management alone. The authors argue that wildfire, landslide, and flood hazards in mountain catchments are interconnected—burned soils shed water faster, lowering the rainfall threshold for slope failure—so a validated fire risk map doubles as an input to multi-hazard planning. The study is candid about its limits: Random Forest remains a black box without regression coefficients or confidence intervals, the static NDVI and land cover layers cannot capture landscape change across the 13-year fire record, and coarse-resolution inputs plus the absence of ground-verified incident data may blur spatial precision. Future work, the team suggests, should apply explainable AI techniques such as SHAP values, incorporate ensemble methods and higher-resolution data, and validate predictions through field surveys and community-based participatory mapping. For now, the map gives district and municipal authorities in Rasuwa something they have never had before: a locally calibrated, statistically validated picture of where the next fire is most likely to start, and where early warning systems and fuel reduction work should begin first.

Subject of Research: Machine learning-based forest fire risk mapping using earth observation data in the high-mountain district of Rasuwa, Nepal

Article Title: Integrating earth observation and machine learning for forest fire risk assessment in Rasuwa district, Nepal

Article References: Integrating earth observation and machine learning for forest fire risk assessment in Rasuwa district, Nepal. (n.d.). https://doi.org/10.1007/s44415-026-00138-0

Image Credits: AI Generated

DOI: 10.1007/s44415-026-00138-0

Keywords: forest fire risk, Random Forest, machine learning, remote sensing, Nepal, Himalaya, GIS, VIIRS, NDVI, Langtang National Park, climate change, risk mapping