Application of learning vector quantization and different machine learning techniques to assessing forest fire...

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning Applications

Background:

• Forest fires pose significant threats to ecosystems and human settlements.
• Accurate forest-fire susceptibility (FFS) mapping is crucial for effective land management and disaster preparedness.
• Geographic Information System (GIS) and machine learning offer powerful tools for spatial risk assessment.

Purpose of the Study:

• To assess forest-fire susceptibility (FFS) in Fars Province, Iran.
• To compare the performance of three GIS-based machine learning algorithms: Boosted Regression Tree (BRT), General Linear Model (GLM), and Mixture Discriminant Analysis (MDA).
• To identify key factors influencing forest fire occurrence.

Main Methods:

• Utilized a database of 358 historical forest fire locations and ten influencing factors (e.g., elevation, rainfall, land use).
• Employed GIS to integrate spatial data and machine learning algorithms (BRT, GLM, MDA) for FFS mapping.
• Validated model performance using ROC curves, accuracy metrics, and 4-fold cross-validation.

Main Results:

• BRT and MDA models demonstrated higher prediction accuracy (AUC ~88% and ~86%) compared to GLM (AUC ~86% and ~82%) for both training and validation datasets.
• The 4-fold cross-validation confirmed the superior performance of BRT and MDA.
• Land use, annual mean rainfall, and slope angle were identified as the most significant determinants of FFS.

Conclusions:

• BRT and MDA models are suitable and accurate for FFS mapping in Fars Province.
• The generated FFS maps can aid in forest resource management and maintaining ecological balance.
• Machine learning approaches provide reliable tools for spatial forest fire risk assessment.