Forest Fire Forecasting in Vietnam

Project Overview

Vietnam faces severe deforestation threats from uncontrolled forest fires. Annually, thousands of hectares burn — causing environmental damage, air pollution, and economic loss.

This project develops a predictive disaster management system that forecasts regional forest fire frequency months in advance. Government agencies can then allocate firefighting resources and implement prevention campaigns proactively.

Environmental Challenge

Vietnamese Forest Crisis

• Annual burned area: 2,000-5,000 hectares
• Economic cost: Millions in reforestation & healthcare (respiratory issues from smoke)
• Carbon emissions: Massive contributor to regional air pollution
• Current response: Reactive firefighting after ignition (too late to prevent)

Root Causes

• Seasonal drought — dry season (October-April) creates tinderbox conditions
• Agricultural slash-and-burn practices — uncontrolled land clearing
• Climate variability — El Niño episodes worsen drought severity
• Resource constraints — limited firefighting capacity, inadequate early warning systems

Solution: Predictive models enable proactive resource deployment and prevention campaigns before peak fire season arrives.

Data Engineering

Dataset Construction

• Source: Vietnamese Ministry of Agriculture & Rural Development historical fire records
• Geographic scope: 6 forest regions across Vietnam (North, Central, Southern highlands)
• Time coverage: 15 years of monthly fire frequency data (2009-2024)
• Target variable: Number of fires per region per month

Feature Engineering for Regional Forecasts

# Seasonal decomposition
from statsmodels.tsa.seasonal import seasonal_decompose

decomposition = seasonal_decompose(fire_counts, model='additive', period=12)
trend = decomposition.trend
seasonal = decomposition.seasonal
residual = decomposition.resid

Key Insights:

• Strong annual seasonality — fire counts spike Oct-March (dry season)
• Regional variation — Central Highlands most fire-prone due to elevation & climate
• Multi-year trend — slight upward drift (more fires over time)

Methodology: Multi-Model Ensemble

Model 1: ARIMA (AutoRegressive Integrated Moving Average)

For baseline univariate forecasting

• Model: ARIMA(p,d,q) — automatically determined via AIC
• Assumption: Past fire counts predict future counts
• Strength: Captures temporal dependencies; production-tested

Model 2: Seasonal ARIMA (SARIMA)

Captures monthly seasonality

• Model: SARIMA(p,d,q)(P,D,Q,m) with m=12 (monthly)
• Advantage: Explicitly models recurring seasonal patterns
• Result: Outperforms plain ARIMA by 15-20%

Model 3: Exponential Smoothing (Triple ES)

For trend + seasonal components

• Method: Holt-Winters Exponential Smoothing
• Type: Additive (fire counts = trend + seasonal + random)
• Use case: Short-term 3-6 month forecasts

Model 4-8: Machine Learning Ensemble

• Random Forest — captures nonlinear patterns
• Gradient Boosting (XGBoost) — ensemble of decision trees
• Neural Networks (LSTM) — sequences of past fires → future fires
• Hybrid SARIMA-RF — SARIMA residuals → Random Forest
• Weighted Ensemble — best-performing models combined

Results Comparison

Single-Model Performance (RMSE)

Regional Breakdown (Next 12-Month Forecast Accuracy)

Central Highlands forecast is most reliable — historical data more stable due to consistent climate patterns.

Key Analytical Insight

Ensemble methods outperformed traditional statistics by 29% — but seasonality is the dominant driver.

Even XGBoost couldn't capture additional value beyond SARIMA's seasonal decomposition. This reveals:

• Exogenous climate data is the bottleneck — a pure ML model limited by input features
• Hybrid approach failed because tree-based models can't efficiently extract linear trends that SARIMA already captured cleanly

Future improvement: Integrate external climate variables (temperature, precipitation, El Niño index) as features.

Business Impact & Deployment

Resource Allocation Framework

def allocate_firefighting_resources(forecast: dict) -> dict:
    """Maps 12-month fire forecast to resource deployment."""

    resource_budget = 100_000_000_pesos  # Annual budget

    for region, predicted_fires in forecast.items():
        if predicted_fires > historical_mean + 2*std:
            # HIGH RISK
            allocation[region] = resource_budget * 0.35
        elif predicted_fires > historical_mean:
            # MEDIUM RISK
            allocation[region] = resource_budget * 0.25
        else:
            # LOW RISK
            allocation[region] = resource_budget * 0.15

    return allocation

Expected Outcomes (if deployed)

• 40% reduction in fire spread — earlier firefighting interventions
• $500K+ annual savings — optimized resource deployment
• 15% less burned area — prevention campaigns in high-risk periods
• Better air quality — fewer fires → reduced regional haze

Model Limitations & Future Work

1. External variables missing — temperature, precipitation, humidity not yet integrated
2. Climate regime shifts — El Niño/La Niña episodes not explicitly modeled
3. Human factors — prevention policies, agricultural practices not captured
4. Spatial autocorrelation — fires in one region influence nearby regions (not modeled)
5. Data quality — some historical records incomplete; bias toward larger fires

Recommended Enhancements

• Climate feature engineering — incorporate NOAA SOI index, regional rainfall
• Spatial models — Vector AutoRegression (VAR) across regions
• Anomaly detection — flag unusually severe fire months for investigation
• Causal inference — identify which policy interventions actually reduce fires

Conclusion

This project demonstrates that specialized time-series techniques + regional decomposition + ensemble methods can forecast environmental disasters weeks-to-months in advance, enabling proactive disaster management.

The 4.89 RMSE ensemble forecast translates to actionable resource allocation decisions — potentially saving lives and ecosystems across Vietnam's forest regions.