Climate change drives earlier wildre season onset in California

12 Wildfires in California have become more frequent in recent decades, with increasingly 13 devastating impacts. The fire season is also lengthening, with an earlier onset. This trend has 14 been hypothesized to be driven by climate change, but it has yet to be quantitatively attributed to 15 climate drivers. Using a comprehensive fire occurrence dataset, we analyze fire season onset and 16 climate controls on its variability and change during 1992-2018 in California’s forested 17 ecoregions. Onset shows an advancing trend, by 14 days in Northern California and 17 days in 18 Southern California. In Southern California, this trend is dominated by decreasing winter 19 precipitation, possibly natural in origin. By contrast, in Northern California, the largest 20 contributor to the advancing onset is a springtime temperature increase, driven to a large degree 21 by climate change. No matter what the dominant contributor to the trend is, the influence of 22

Fire season onset is a fire behavior characteristic with practical applications in fire risk 42 outlooks (NICC, 2021). An earlier than usual onset is an indicator of a potentially longer fire 43 season, and therefore more fire risk with a longer mean burning time (Westerling, 2016). 44 Additionally, when the onset is earlier, the environment has more potential to become primed for 45 subsequent large fires (Khorshidi et al., 2020). Onset has been widely identified as when fire 46 weather conditions first surpass a selected threshold (e.g., Jolly et al., 2015). For instance, 47 examining Canadian forests, Wotton & Flannigan (1993) defined onset as when maximum daily 48 temperature starts to exceed 12 o C for three consecutive days. Such an emphasis on temperature 49 in defining onset has naturally led to the conclusion that climate change has the potential to 50 advance onset in many regions (e.g., Wotton & Flannigan, 1993;Strydom & Savage, 2017), 51 including California (Abatzoglou & Kolden, 2011). While fire weather is indicative of the fire 52 danger, fire occurrence also depends on the fuel availability and ignition (Moritz et al., 2012). 53 Additionally, widely-used fire weather metrics may not represent complex hydrological 54 processes such as snow influence (Abatzoglou & Williams, 2016), which also influences aridity 55 and hence fuel flammability. This highlights the importance of investigating fire season onset 56 using fire occurrence data to develop a robust understanding of its drivers, thereby identifying 57 the influence of climate change (Williamson et al., 2016).

58
Fire-occurrence-based analysis of fire season onset has previously been constrained by 59 the limitation of fire occurrence data (Williamson et al., 2016). The question of how much 60 climate change and natural variability have contributed to the changes in historical fire season 61 onset in California has yet to be answered (Westerling et al., 2006;Westerling, 2016 (Jin et al., 2015). Offshore-wind-dominated large fires in 68 Southern California, in particular, are limited by human ignitions (Keeley et al., 2021). This 69 stresses the need for assessing the onset using comprehensive fire records with a wide spectrum 70 of fire sizes and an objective definition of onset.

71
In this study, we use an extensive record of fire occurrence data for 1992-2018 for 72 ecoregions in Southern and Northern California ( Supplementary Fig. 1

80
Fire occurrence record-based fire season onset 81 The probability distributions of the discovery dates of fires during each calendar month 82 (Fig. 1a,d) show that each month from October to May produces fewer than 10% of total annual 83 fires in both domains. The number of fires increases in summer. The peak occurs in July, with 84 18% and 26% of the annual fires in Southern California and Northern California, respectively. 85 We define the mean discovery date of the year (i.e., Julian date) of all fires, regardless of their 86 size or other characteristics, as the fire season onset (hereafter, onset). The mean onset from 87 1992-2018 is 205 (24th July) and 210 (29th July) for Southern and Northern California, between onset and size is beyond the scope of this study. However, this significant relationship, 103 along with the variability of onset associated with dry and wet years, suggests the onset metric is 104 a predictor of critical fire behavior later in the season. Furthermore, from June to July (i.e., when 105 the onset approaches), cumulative burned area surges from 6.2% to 24.7% of the annual total in 106 Southern California, and from 2.8% to 10% in Northern California ( Supplementary Fig. 3). This 107 steep increase in burned area during onset illustrates that our definition of onset is not an abstract 108 quantity but rather represents a step change in the development of the fire season. The fact that 109 onset is swiftly followed by a rapid increase in burned area is further evidence that a proper  Northern California has a larger snowpack (Minnich, 2018), which could explain the above 135 differing seasonal influences of precipitation and temperature for two domains. Based on these 136 results, we hypothesize that through surface and subsurface moisture buffers, winter to spring 137 precipitation influence is being carried over to summer in Northern California, influencing onset.

138
To test the mechanism hypothesized above, through which winter-spring precipitation 139 and temperature influence the summer fire season onset, we look at soil moisture profiles. Soil  Supplementary Fig. 4). The change of R 2 profile with depth for a given month, however, 152 may reflect the relationship between moisture at that level and plant water uptake, which may 153 depend on various vegetation traits and environmental factors (e.g., Fellows and Goulden, 2017). 154 Next, we test this hypothesis for the two domains by looking at results from soil moisture at 155 different depths.

156
In Northern California, when moving to deeper soil layers, the peak R 2 of soil moisture 157 shifts to later months of the year (Fig. 3d-f). The deepest layer considered here (100-200m) 158 shows the strongest relationship with onset (R 2 =0.74) in JJA, i.e., coinciding with the onset. In Northern California, soil moisture at 100-200cm varies considerably over the year, e.g., the 160 correlation between NDJ and JJA soil moisture is 0.55 ( Supplementary Fig. 5). This likely 161 results from the large contribution of snow melt and increasing evaporative demand when 162 proceeding from winter to summer ( Supplementary Fig. 4), temperature-influenced processes  In both domains, the shift in the peak R 2 with depth is approximately consistent with 177 volumetric soil moisture (i.e., volume of water per unit volume of soil) declining below a certain 178 threshold in each soil layer ( Supplementary Fig. 4). This is consistent with plant water stress 179 being triggered by soil moisture decreasing below a threshold corresponds to plant physiological

197
The linear model of onset using Tv, Pv and CC adequately captures the observed onset 198 (Fig. 4a,b). Inspecting the contribution of individual components allows us to understand their 199 roles during extreme events (Fig. 4d,e). For instance, during the 2012-2015 extreme drought, the 200 early onset in Southern California (Fig. 4a) is linked mainly to dry conditions (i.e., precipitation 201 deficit) (Fig. 4d), whereas in Northern California (Fig. 4b), onset is regulated not only by the 202 precipitation deficit, but also the warm conditions produced jointly by temperature variability 203 (Tv) and climate change (CC). All three components contribute similarly to the onset anomaly 204 (Fig. 4e). Decomposition of the linear trend of observed onset into individual components (Fig.   205 4c) shows that in Southern California, the dominating component is Pv, which accounts for -13 206 days (90% CI -25 to -3 days), with a moderate contribution from CC, which accounts for -3 days 207 (90% CI -12 to 4 days), and negligible impact from Tv (90% CI -4 to 4 days). For Northern 208 California, CC is the largest contributor, accounting for -7 days (90% CI -17 to 2 days), with 209 smaller contributions from Pv (-4 days, 90% CI -13 to 2 days) and Tv (-2 days, 90% CI -9 to 4 210 days). We also find that in both regions, the contribution from CC to the onset trend during     320 To estimate the influence of climate variability and climate change, we first develop a 321 simple causal network (Supplementary Fig. 6). We take temperature and precipitation as the influence of the variable with the highest correlation (say Z). This is done by calculating the 332 partial correlation. If a variable X has a statistically significant (p<0.01) partial correlation with 333 onset, it is declared that X has an influence on onset independent of Z. If the partial correlation is 334 not significant, the variable is removed. This way we objectively obtain a smaller subset of P0 335 (say P1). This is repeated by conditioning on the variable with the next highest absolute 2). Repeating the above analysis using vapor pressure deficit, a fire weather metric that is highly  With the above selected drivers of onset, we proceed to quantify their influence on onset. 343 First, we assume that climate change is influencing onset through temperature and precipitation 344 ( Supplementary Fig. 6). To isolate this influence from natural climate variability, we use a global  where α, β and γ are the regression coefficients, λ is the intercept, and ε is the residual.

Estimation of the influence of climate variability and climate change
It should be noted that, we do not assess the uncertainty arising from the method used to   Mann-Kendall trend test (Hamed & Rao, 1998).