Figure 1 . (a) Arctic tundra region in Alaska as defined by CAVM. (b) Number of fire events detected with MCD14ML data from 2001 to 2019.
3.2 Empirical modeling performances
Three groups of independent variables show very strong correlations, including the vegetation indices representing fuel moisture state, the CFFWIS components representing fire weather conditions, and the topographic features, with Pearson’s r above 0.8 (Figure S3). Since all vegetation indices were highly correlated with Pearson’sr above 0.95, we only adopted NDII6 to estimate fuel moisture state for further modeling efforts. Strong correlations were also found between the fire behavior indices (ISI and BUI) and fuel moisture codes (FFMC and DMC) of the CFFWIS. Since we did not focus on fire propagation, we only selected the three fuel moisture codes to represent fire weather conditions. Although the near-surface weather variables show moderate correlations with the CFFWIS components, they were included to account for meteorological conditions irrespective of fuels. Additionally, slope and roughness were removed for modeling due to their strong correlations with elevation.
Both the “Current-day model” and “Previous-day model” developed with the RF classification algorithm have shown a strong capability in predicting the fire occurrence probability in the tundra. The overall out-of-bag (OOB) error rate of the “Current-day model” is 6.03%, with the overall accuracy reaching 93.97% (Table S3). The “Previous-day model” shows slightly lower modeling performance, with an overall OOB error rate of 8.75% and an accuracy of 91.25%. Validation performed against the reserved dataset shows that both models can reflect (with the “Current-day model”) and forecast (with the “Previous-day model”) fire occurrence probability, as indicated by the Receiver Operating Characteristic (ROC) curves (Figure S4). The Area Under the Curve (AUC) values reached 0.97 and 0.96 for the “Current-day model” and the “Previous-day model”, respectively.
3.3 Environmental factors driving tundra fire occurrence
CG lightning probability was identified as the most important variable in both the “Current-day model” and “Previous-day model” for predicting tundra fire occurrence, with Mean Decrease in Accuracy (MDA) of 50.06% and 34.58%, respectively (Figure 2 a-b). A significant positive relationship was confirmed between CG lightning and fire occurrence via logistic regression models (p < 0.001; Table 1), suggesting that regions with larger lightning probability are likely to experience higher fire risks. On fire-occurrence days, the lightning probability of the “Fire” events were higher than 0.50 on average across the tundra region and reached over 0.62 in the North Slope and Southwest Alaska (Figure 2 c). In contrast, the lightning probability was below 0.15 on average when no fire occurred. Similarly, on the previous days of fire occurrence, though lower than that on fire-occurrence days, the lighting probability of the “Fire” events, was significantly higher (~0.48) than that of the “No fire” events (< 0.12) on average (Figure 2 d).