Comparing four regression techniques to explore factors governing the number of forest fires in Southeast, China

• Main Authors: Qianqian Cao, Lianjun Zhang, Zhangwen Su, Guangyu Wang, Shuaichao Sun, Futao Guo
• Format: Article
• Language: English
• Published: Taylor & Francis Group 2021-01-01
• Series: Geomatics, Natural Hazards & Risk
• Subjects: geographically weighted poisson regression; geographically weighted negative binominal regression; forest fire count; over-dispersion; spatial autocorrelation and heterogeneity
• Online Access: http://dx.doi.org/10.1080/19475705.2021.1884609
• ISSN: 1947-5705; 1947-5713

Four regression techniques, including two global models (i.e., Poisson and negative binominal) and two geographically weighted regression (GWR) models (i.e., geographically weighted Poisson regression (GWPR) and geographically weighted negative binominal regression (GWNBR)) were used to explore which was the most suitable method for predicting the number of forest fires and to investigate the spatially varying relationships between forest fires and environmental factors in Fujian province, in the Southeast of China. Our results showed that the GWR models fitted the fire count data better than the global models, and yielded more realistic spatial distributions of model predictions. Particularly, GWNBR was superior for addressing overdispersion in the fire count data because it estimated the dispersion parameter at a local level. Additionally, our study indicated that more forest fires occurred in areas of lower elevation, flatter terrain, and higher population density. The global models showed that precipitation had positive impacts on fire occurrence in the study area. In contrast, the GWR models revealed that precipitation was positively related to the forest fires in the western regions of Fujian, but negatively related in the eastern coastal regions. Our study could provide better insight into forest fire management based on local environmental characteristics.