A spatial weight sampling method integrating the spatiotemporal pattern enhances the understanding of the occurrence mechanism of wildfires in the southwestern mountains of China

Abstract

Wildfire risk assessment is crucial for the protection of mountainous ecosystems. While data-driven machine learning algorithms have advanced wildfire modeling, the quality of non-fire sample selection remains a significant limitation for model performance. This study proposes a spatial weight-based sampling method to improve non-fire sample quality, enhance model precision, and provide deeper insights into wildfire occurrence mechanisms. The study area is Yunnan Province, situated in southwest China, is dominated by mountainous landscapes and exhibits a significant susceptibility to wildfires. A spatial weighting model was developed using historical wildfire data (2011–2020), and threshold values were applied to divide different non-fire sample collection ranges. Four machine learning algorithms (KNN, SVM, RF, and DNN) were employed to construct prediction models, and their performance and spatial consistency were analyzed to determine the optimal sampling range and best-performing model. The results showed that (1) the optimal sampling weight range was 0–0.4, with RF achieving the best performance (AUC = 0.90). (2) The key factors influencing wildfire occurrence are population density and elevation. (3) The spatial weight-based sampling method demonstrated significant advantages in prediction accuracy and spatial consistency. This approach offers a novel framework for the collection of non-fire samples, and an improved understanding of wildfire occurrence mechanisms, contributing to more effective risk mitigation strategies.