Assessment of the effects of different variable weights on wildfire susceptibility

In this study, wildfire susceptibility is mapped using various multi-criteria decision analysis techniques (AHP, SAW and VIKOR) and machine learning algorithms (MaxEnt and logistic regression) to reveal the response of models for wildfires. In this study, it is suggested that using natural weights generated by machine learning algorithms instead of artificial weights in MCDA methods can increase the reliability of susceptibility maps because wildfires have very close relationship with climatic, topographic and environmental variables. The contribution rates (natural weights) were obtained using machine learning algorithms and incorporated into MCDA methods to make the spatial relationships between variables more obvious. As a result, eight susceptibility maps were generated using MCDA methods, MaxEnt and logistic regression algorithms. Correlation analysis showed that using natural weights instead of artificial weights increased the correlation between MCDA methods and machine learning algorithms. Each correlation value increased by 10% on average and the highest increase was determined between VIKOR and logistic regression from 0.6286 to 0.7580 when natural weights were used. In addition, 1035 existing wildfire locations were used to evaluate the reliability of generated maps. The results showed that the average risk values of 1035 wildfire locations increased from 6.04 to 7.23 using AHP, from 0.66 to 0.79 using SAW and from 0.35 to 0.25 using the VIKOR method. This indicates a significant increase in the accuracy and reliability of susceptibility maps produced when natural weights determined by machine learning algorithms are used in MCDA methods.