Reliability evaluation of pseudo-samples from model-free and model-based approaches in wildfire susceptibility mapping

Given the profound impact of wildfires on human society and the natural environment, as well as the challenges faced by traditional supervised learning methods when sample sizes are insufficient, this study aims to evaluate the reliability of model-based and model-free semi-supervised learning (SSL) approaches in generating pseudo-samples for wildfire susceptibility mapping. The key distinction between these SSL approaches lie in their reliance on specific predictive models, such as random forest, support vector machine, or logistic regression, during pseudo-sample generation. In this study, we employed two semi-supervised learning methods: self-training (representing the model-based approach) and label propagation (representing the model-free approach) to create wildfire susceptibility maps in Wanzai County, Jiangxi Province, China. The results show that self-training combined with random forest exhibited optimal performance in the reliability evaluation of pseudo-sample quality, achieving an overall accuracy (OA) of 80.1%. In addition, label propagation also demonstrated high reliability with 77.5% of the OA. Further indirect reliability evaluation confirmed that integrating pseudo-samples into the original sample set enhances the accuracy of wildfire susceptibility mapping, particularly when the number of pseudo-samples is limited to 1,000 or fewer. Moreover, this study explores the motivations behind the reliability evaluation of pseudo-samples, potential efficacy mechanisms, and the applicability.