Multivariate Fire Risk Models Using Sample-Size Reduction and Copula Regression in Kalimantan, Indonesia

Abstract

Forest fires have become a national issue every year and get serious attention from the government and researchers, especially in Kalimantan. The copula-based joint distribution can construct a fire risk model to improve the early warning system of forest fires. This study aims to model and analyze the copula-based joint distribution between climate conditions and hotspots in Kalimantan. We constructed the bivariate joint distributions between climate conditions, either total precipitation or dry spells, and hotspots with sample size reduced by ENSO conditions, i.e., La Nina, normal, and El Nino. From the joint distribution, fire risk models are calculated using conditional probability and copula regression. The results show that the relationship between climate conditions and hotspots in La Nina and normal ENSO conditions have an upper tail dependence but no lower tail dependence. Meanwhile, the relationship has both upper and lower tail dependences during El Nino. There is an outlier in normal ENSO conditions with more hotspots than normally, i.e., in September 2019. The probability is very low during normal ENSO conditions, i.e., less than 2%. The only relatively high probability is during El Nino, i.e., more than 10%. Moreover, the copula regression models show that the model given specific dry spells is better than that given specific total precipitation as climate condition. The copula regression for hotspots given specific total precipitation and ENSO conditions has the RMSE value of 1339 hotspots and the R2 value of 60.70%. Meanwhile, the copula regression for hotspots given specific dry spells and ENSO conditions has the RMSE value of 1185 hotspots and the R2 value of 69.21%.