Wildfire Forecasting & Climate Projections
Research Associate Project | Central Department of Hydrology and Meteorology, Tribhuvan University
July 2024 – December 2025 | Prediction & Projection Study
Project Overview
This comprehensive research project addresses wildfire risk in Nepal through two integrated components: (1) short-term forecasting of wildfire occurrence using machine learning, and (2) long-term projections of wildfire risk under future climate scenarios. The project combines meteorological data, environmental factors, and advanced machine learning with state-of-the-art climate projections.
Wildfires pose increasing threats to Nepal's ecosystems, air quality, and economic sectors. This research provides evidence-based insights for disaster risk reduction and climate adaptation planning.
Interactive Wildfire Risk Visualization
Loading map...
Short-Term Wildfire Forecasting
Machine Learning Approach
Developed Random Forest machine learning models to predict wildfire occurrence at sub-seasonal timescales. Random Forests are ensemble learning methods that excel at capturing non-linear relationships and interactions between predictor variables.
Meteorological Predictors
The models integrate multiple meteorological variables that drive wildfire risk:
- Temperature: Higher temperatures increase vegetation drying and flammability
- Relative Humidity: Low humidity strongly correlates with fire danger
- Wind Speed: Wind influences fire spread and intensity
- Precipitation: Recent rainfall reduces fuel dryness and fire risk
- Evapotranspiration: Indicates moisture stress in vegetation
Environmental Predictors
- Elevation & Terrain: Slope and aspect influence fire behavior
- Land Use / Land Cover: Forest type and density affect flammability
- Vegetation Index (NDVI): Remote sensing-based vegetation condition
- Population Density: Human activities and ignition sources
- Proximity to Roads: Access routes and human presence
Data Sources
- GFS Forecast Data: Global Forecast System weather predictions (nowcasts to 10-day)
- Observational Networks: Surface weather stations across Nepal
- Remote Sensing: MODIS, Sentinel for land cover and vegetation indices
- Historical Fire Records: Documented wildfire events for training
Model Performance & Validation
Models validated using standard classification metrics including:
- Precision and Recall for fire occurrence prediction
- ROC-AUC (Receiver Operating Characteristic - Area Under Curve)
- Feature Importance ranking for interpretability
- Cross-validation to assess generalization capability
Future Wildfire Projections
Climate Projection Framework
Projected end-of-century (2070-2100) wildfire risk using bias-corrected climate model outputs from the latest Coupled Model Intercomparison Project (CMIP6). This approach combines machine learning with climate science to assess how wildfire risk may change under different emission scenarios.
Emission Scenarios (SSP Pathways)
Projections created under two contrasting Shared Socioeconomic Pathways (SSP):
SSP2-4.5 (Moderate Emissions)
Intermediate emissions scenario assuming continued global development with some climate mitigation efforts. Global warming reaches approximately 2°C by 2100 relative to pre-industrial levels.
SSP5-8.5 (High Emissions)
High emissions scenario assuming fossil fuel-intensive development with limited climate mitigation. Global warming reaches approximately 4.5°C by 2100. Represents worst-case scenario for wildfire risk escalation.
CMIP6 Multi-Model Ensemble
Used ensemble projections from multiple climate models to capture uncertainty:
- Model Ensemble: 10+ global climate models from CMIP6
- Bias Correction: Quantile-based bias correction to minimize systematic errors
- Downscaling: Spatial and temporal downscaling to regional resolution (~3-4 km)
- Ensemble Mean: Multi-model mean used for robust projections
- Uncertainty Quantification: Spread across models indicates confidence level
Integration with ML Models
Downscaled climate projections serve as inputs to the Random Forest models, enabling translation of future meteorological conditions into wildfire risk probabilities. This integrated approach captures both climate change effects and non-linear fire dynamics.
Key Findings
Forecast Skill
Random Forest models demonstrate significant forecast skill for predicting wildfire occurrence at sub-seasonal timescales using GFS meteorological data.
Key Predictors
Temperature, humidity, and recent precipitation are the dominant drivers of wildfire occurrence. Vegetation condition (NDVI) also plays important role.
Future Risk Escalation
Projections indicate increasing wildfire risk across Nepal by end of century, with stronger increases under high emission scenario (SSP5-8.5).
Regional Variations
Risk increases vary spatially, with highest increases in lower elevation forests and mid-hill regions where human-fire interactions are significant.
Seasonal Patterns
Spring (pre-monsoon) remains the primary wildfire season, with potential for extension of high-risk periods into early monsoon under warmer climate.
Compound Hazards
Increasing wildfire risk compounds other climate-related hazards including floods, landslides, and air pollution, affecting multiple sectors and communities.
Technologies & Tools
Machine Learning
- Scikit-learn (Random Forest)
- Python
- Feature Engineering
- Model Validation
Climate Data
- CMIP6 Projections
- GFS Forecast Data
- Bias Correction
- Downscaling Methods
Remote Sensing
- MODIS Data
- Sentinel Imagery
- Google Earth Engine
- Vegetation Indices (NDVI)
Data Processing
- Python (xarray, Pandas)
- NetCDF Analysis
- GIS Analysis
- Statistical Methods
Related Publications & Reports
Peer-Reviewed Articles
- Kuikel, S., Paudel, H.K., et al. (2025). "Impact of wildfire smoke on air pollution-related premature mortality in rapidly growing Kathmandu Valley."Atmospheric Environment X, 26, 100334. https://doi.org/10.1016/j.aeaoa.2025.100334
Research Reports
- Pokharel, B., Marahatta, S., Giri, S., Pradhananga, D., Kuikel, S., Joshi, K.P., Kuinkel, D., Sapkota, S., & Paudel, H.K. (2025).Wildfires in Nepal: Historical Trends, Current Status, Future Projections, Impacts, and Public Perception. Central Department of Hydrology and Meteorology, Tribhuvan University. ISBN: 978-9937-1-9441-9
Research Impact & Applications
- Disaster Risk Reduction: Provides early warning capability for wildfire occurrence at sub-seasonal timescales to support emergency response.
- Climate Adaptation: Future projections inform long-term adaptation strategies for forestry, agriculture, and disaster management sectors.
- Air Quality Management: Wildfire projections support air quality forecasting and public health protection, especially for vulnerable populations.
- Policy Support: Results support evidence-based policy formulation for climate action and disaster risk reduction at national and regional levels.
- Capacity Building: Framework transferable to other South Asian countries for regional wildfire risk assessment.
- Integrated Risk Assessment: Contributes to understanding of compound climate hazards and cascading impacts on communities.
Future Work & Extensions
- Incorporate ensemble seasonal predictions for enhanced forecast skill at longer lead times
- Develop probabilistic wildfire forecasting systems with explicit uncertainty quantification
- Integrate socioeconomic factors and land-use change scenarios into future risk projections
- Implement operational early warning system integrated with Nepal's disaster management platforms
- Assess cascading impacts of wildfire on water resources, agriculture, and air quality