Modeling the spatial patterns of human wildfire ignition in Yunnan province, China

Despite wildfire being an important regulator of dryland ecosystems, uncontrolled wildfire can be harmful to both forest ecosystems and human society, and wildfire prevention and control continue to raise worldwide concern. Wildfire management depends on knowledge of wildfire ignitions, both for cause and location. The regimes and factors influencing wildfire ignition have been studied at length. Humans have a profound effect on fire regimes and human activity is responsible for igniting the largest number of fires in our study area. Understanding the spatial patterns of ignitions is foremost to achieving efficiency in wildfire prevention. Previous studies mainly concentrate on overall wildfire risk integrating numerous factors simultaneously, yet the importance of human factors on ignition has not received much attention. In this study, we mapped human accessibility to explore the influence of human activity on wildfire ignition in a simple and straightforward way. A Bayesian weights-of-evidence (WofE) method was developed based on fire hotspots in China's Yunnan province extracted from satellite images and verified as known wildfires for the period 2007-2013. We considered a set of factors that impact fire ignition as associated with human accessibility: the locations of settlements, roads, water and farmland susceptible to human wildfire ignition. Known points of likely wildfire ignition were selected as training samples and all suspected thematic maps of the factors were taken as explanatory layers. Next, the weights of each layer in terms of its explanatory power were computed and used to generate evidence based on a threshold to pass a statistical test. The conditional independence (CI) of each layer was checked with the Agterberg-Cheng test. Finally, the posterior probability was calculated and its precision validated using samples of both presence and absence by withheld validation data. A comparison of WofE models was made to test the predictability. Results show proximity to villages, roads and farmland are strongly associated with human wildfire ignition and that wildfire more often occurs at an intermediate distance from high-density human activity. The WofE method proved more powerful than logistic regression, improving predictive accuracy by 10% and was more straightforward in presenting the association of dependence and independence. In addition, WofE with 1000 m buffer bands is more robust in predicting human wildfire ignition risk than binary or 100 m buffers for the ecoregion studied. Our results are significant for advising practical wildfire management and resource allocation, evaluation of human ignition control and also provides a foundation for future efforts toward integrated wildfire prediction.