A novel optimized repeatedly random undersampling for selecting negative samples: A case study in an SVM-based forest...

Area of Science:

• Geospatial science
• Machine learning
• Environmental risk assessment

Background:

• Accurate negative sample selection is crucial for machine learning-based natural hazard assessments.
• Existing methods like Single Random Sampling (SRS) and Repeatedly Random Undersampling (RRU) have limitations in ensuring optimal classifier performance.

Purpose of the Study:

• To propose an optimized Repeatedly Random Undersampling (RRU) method for improved negative sample selection.
• To enhance the classification performance of machine learning models in natural hazard assessment.
• To generate a more accurate forest fire susceptibility map for decision support.

Main Methods:

• Developed an optimized RRU method to identify the actual most accurate classifier (MAC).
• Utilized Support Vector Machine (SVM) as the analysis method.
• Employed Genetic Algorithm to optimize SVM parameters for forest fire susceptibility assessment in Huichang County, China.

Main Results:

• The optimized RRU successfully identified an actual MAC with superior classification performance compared to the standard RRU.
• The forest fire susceptibility map generated using the actual MAC aligns with factual observations.
• The optimized RRU demonstrated enhanced negative sample selection, improving assessment reliability and accuracy.

Conclusions:

• The optimized RRU is a more effective approach for negative sample selection in natural hazard studies.
• The developed method significantly improves the accuracy and reliability of forest fire susceptibility assessments.
• The findings provide valuable decision support for local governments in mitigating forest fire risks.