Figure 2 . Variable importance rankings of (a) the “Current-day model” and (b) the “Previous-day model”. Boxplots of CG lightning probability for the “Fire” and “No fire” events in the three tundra regions on (c) fire-occurrence days and (b) the previous days before occurrence.
WRF-simulated near-surface meteorological variables and fuel moisture codes, particularly air temperature, RH, and DC, were also found important in modeling tundra fire occurrences, as indicated by MDA values from the RF models (Figure 2 a-b). Specifically, higher air temperature and drier fuels could contribute to increases in fire occurrence probability, according to the significantly positive relationships between temperature and DC with fire occurrence (p < 0.05; Table 1). The mean air temperature was significantly higher in most tundra regions when fires occurred, while RH was significantly lower (Table S4). On fire-occurrence days, the air temperature of the regions with fires can reach 24.8◦C and 23.5◦C in Southwest Alaska and the North Slope on average, respectively. In comparison, regions with no fires were much cooler, with 18.4◦C and 16.5◦C, respectively (Figure 3 a). As expected, drier conditions were also likely to support fire occurrence. The RH values of “Fire” events were about 9.6% lower than those of “No fire” events in these two tundra regions on average (Figure 3 b). In addition, all fire weather indices were significantly higher on fire days in North Slope and Southwest Alaska. Though Alaskan tundra is not a moisture-limited ecosystem, surface vegetation fuels can dry out rapidly to support burnings, with FFMC reaching above 80 across the tundra regions on the fire-occurrence days (Figure 3 c). Moreover, the significantly negative relationships between NDII6 and fire occurrences in both logistic regression models indicated that drier fuels support burnings in the tundra (p < 0.05; Table 1). Mean values of the vegetation indices related to fuel moisture state were slightly but significantly lower for the “Fire” events (Figure S5; Table S4).
Table 1. Logistic regression results of the two models.