Trends in tree cover change over three decades related to interannual climate variability and wildfire in California

The U.S. State of California has experienced frequent drought events, hotter temperatures and other disruptions to the climate system whose effects on ecosystems have been widely reported in recent decades. Studies primarily confined to specific vegetation communities or species, individual drought incidents, or analysis over a relatively short intervals, has limited our understanding of the broad-scale effects on tree cover and the spatiotemporal variability of effects across broader regions. We focused analysis on multi-annual land cover and land surface change to assess patterns and trends in tree cover loss in tree-dominated Californian ecoregions from 1986 to 2019. The top three years of total tree cover loss for the state were 2018 (1901 km2), 2015 (1556 km2), and 2008 (1549 km2). Overall, annual tree cover loss had upward trends. Tree cover loss rapidly surged later in the study period and was apparently driven by climate stress and wildfires. Underlying geographic variability was apparent in both non-fire and fire-related tree cover loss that sharply increased during hotter multi-year droughts. The increasingly hotter and drier climate conditions were associated with significant increases in fire-induced mortality. Our findings indicate that a possible effect of future hotter and drier climate would lead to further tree cover loss, thereby endangering California’s ecosystem goods and services. Geographic variability in tree cover trends indicates that ecoregion-specific mitigation and adaptation strategies would be useful to conserve the region’s forest resources. Such strategies may benefit from consideration of historical disturbances, ecoregion’s sensitivity to disturbance types, as well as potential ecoregion-specific climate-vegetation-fire feedbacks.


Introduction
Increased warming along with persistent and intense droughts have elevated climate-driving disturbances across the western United States. In particular, the U.S. State of California has been experiencing frequent drought events, hotter temperatures, and other disruptions to the climate system whose effects on ecosystems have been widely reported in the last few decades. Notably, California's drought of 2012-2016 has been considered as one of the most consequential droughts in over a century (Griffin and Anchukaitis 2014, Robeson 2015, Lund et al 2018. This drought affected millions of trees, resulting in widespread crown defoliation, foliage die-back and massive die-offs (Byer and Jin 2017, Paz-Kagan et al 2017, Stephenson et al 2018, Keen et al 2020. Critical habitats of endemic tree species suffered largescale tree mortality during that warmer drought (Das et al 2020, Dwomoh et al 2021. Tubbesing et al (2020) estimate that approximately 1%-5% of the aboveground live tree biomass across California's forested lands died during the drought. Moreover, wildfire activity has been trending upward in the western United States, further presenting increased risk for widespread tree mortality (Westerling 2016, Not subject to copyright in the USA. Contribution of U.S. Geological Survey (USGS) Environ. Res. Lett. 18 (2023) Williams et al 2019, Huang et al 2020, Buechi et al 2021.
Many studies have documented adverse effects on vegetation communities as a result of the changing climate and changing wildfire regimes in California (Williams et al 2019, Huesca et al 2021. Nonetheless, most of the studies were confined to specific vegetation groups or few species, specific drought incidents, or analysis over a relatively short annual time period (Baguskas et al 2014, Taylor et al 2019, Baeza et al 2021, Wayman and Safford 2021, Robbins et al 2022; thus, limiting our understanding of the broadscale effects on tree cover as well as the spatiotemporal variability of effects among ecoregions across California. In this study, we used multiple annual land cover and land surface change products from the U.S. Geological Survey's (USGS) Land Change Monitoring, Assessment and Projection (LCMAP) product suite to assess patterns in tree cover loss in eight tree-dominated ecoregions in California from 1986 to 2019. Furthermore, we combined LCMAP-derived tree cover loss maps with climate and wildfire datasets to assess effects of fire and climate on tree cover loss and identified hotspots of tree cover loss and fireinduced tree cover declines across the study area.
Much of recent wildfire studies in the western United States have focused on patterns of burned area and burn severity at multiple scales (Steel et al 2018, Parks and Abatzoglou 2020, Iglesias et al 2022. Often missing from previous studies are the effects of wildfires on tree cover. Nonetheless, changes in tree cover are not only associated with but are also key indicators of vital ecological processes including loss of treeheld carbon, tree mortality, diminished productivity, soil health, water quality, loss of habitat, and biodiversity. Thus, to expand on spatial and temporal understanding of recent change across California, we evaluated the patterns of tree cover decline by fire which could be interpreted as fire-related tree mortality. Moreover, assessing patterns of tree cover decline and their relationships to interannual climate variability including drought, and related disturbances such as wildfire is important for developing effective mitigation and adaptation strategies in a warming and drying world. We specifically addressed the following research questions: (a) What are the overall temporal trends in annual area of tree cover loss across the study area? (b) Did the patterns of tree cover loss vary across ecoregions during periods of notable drought events in the study period? (c) Which climate variables are most associated with fire-induced tree cover loss for each ecoregion? (d) What are the temporal trends of climate variables most associated with fire-induced tree cover loss for each ecoregion?  1). These areas cover a mosaic of land cover types including forested vegetation, agriculture, and grasslands.
In the last few decades, California was affected by several drought events with considerable effect on ecosystems and adverse economic impacts. The state's multi-year drought of 2012-2016 has been considered as one of the most severe droughts in over a century (Griffin and Anchukaitis 2014, Robeson 2015, Lund et al 2018. During that drought several millions of trees were severely affected leading to substantial dieoff (Byer andJin 2017, Stephenson et al 2018) and degradation of critical habitats (Dwomoh et al 2021).

Land-cover and land-change dataset
In our study, land cover and land change data at an annual frequency provided the temporal specificity to explore the location and timing of change and determine trends. We obtained annual land cover and land surface change products of LCMAP Collection 1.2 from the USGS Earth Resources Observation and Science (EROS) Center. Thirty-meter resolution LCMAP products are based on the USGS implementation of the Continuous Change Detection and Classification (CCDC) algorithm (Zhu and Woodcock 2014) on the Landsat archive (Collection 1) to monitor land-cover and land-cover change (Brown et al 2020, Xian et al 2022. The LCMAP product suite has ten science products (https://doi. org/10.5066/P9W1TO6E) shown in table 1.
More information on these products can be found in Brown et al (2020). In this study we used LCPRI, LCACHG, LCPCONF, SCTIME, and SCMAG (table 1).
In this study, Tree Cover refers to a class in the land cover legend of LCMAP, which is often called 'forest' in other mapping efforts. A land cover is considered a tree cover class if it is a vegetated land containing 10% or greater tree cover. Thus, the tree cover category is one of eight land cover  classes in each annual LCPRI product, whereas the LCACHG provides an indicator of annual land cover change by each of the eight land cover classes. Tree cover loss is specifically identified when pixels in the tree cover class (class 4) in one year change to another land cover type in the following year. In this study, we use the term 'tree cover change in condition' to describe a change in tree cover condition without a land cover conversion (i.e. change in land cover type). Tree cover change in condition can result from harvest thinning, windthrow, insect attacks, moisture stress, wildfire, and other disturbances that reduce tree cover but not to below the 10% threshold used in the LCMAP classification of tree cover class. For each calendar year, we calculated tree cover loss and tree cover change in condition following procedure detailed in supplementary materials (SM) S1.

Fire dataset
We obtained fire perimeter and burn severity data for all large wildfires (>4 km 2 ) from the Monitoring Trends in Burn Severity (MTBS, www.mtbs.gov/) project (Eidenshink et al 2007). Timing of fire occurrence were obtained from the fire ignition and extinction dates in the MTBS fire perimeter data. We separated annual tree cover loss and tree cover change in condition (also referred to as tree cover condition change) into those related to fire and non-fire causes by overlaying with SCTIME and burned area information from the MTBS burned area data. Details of our approach for partitioning tree cover loss into fire and non-fire causes are provided in SM S1.

Climate data
We obtained daily gridded meteorological dataset at 1/24 • spatial resolution (≈4 km) from gridMET (www.climatologylab.org/gridmet.html, Abatzoglou 2013). We considered the following climate variables from the dataset: total precipitation (Precip), minimum temperature (Tmin), maximum temperature (Tmax) and mean vapor pressure deficit (VPDmean). Vapor pressure deficit (VPD) is the difference between the amount of moisture in the air and how much moisture the air can hold when saturated (Seager et al 2015). VPD is a measure of atmospheric aridity and used as an indicator of drought severity in several ecological studies (Restaino et al 2019, Higuera and Abatzoglou 2021, Juang et al 2022. For all variables, we calculated annual and seasonal summaries and standardized climatic anomalies for the water year (October-September), calendar year (Tmin and Tmax only), and the fire season, generally considered as May-October in the western United States (Westerling et al 2003). We calculated standardized climatic anomalies (z) as the annual (X) deviation from the climatological mean (µ ref ) divided by standard deviation of the climatological mean (σ ref ) for the reference period 1986-2015 using equation (1). All climate variables were calculated for each gridMET grid cell and summarized for each ecoregion as well as the entire state of California: (1)

Temporal trends in annual area of tree cover loss
We summarized annual tree cover loss pixels for each ecoregion and the entire state of California for the study period 1986-2019. We calculated the following metrics of tree cover loss for further analyses: total tree cover loss, fire-induced tree cover loss, non-fire tree cover loss, and fire-induced tree cover change in condition. Due to the lack of statistically based reference sample data targeted specifically for our inquiry, our tree cover loss area estimates were based on pixel counts in the LCMAP landcover dataset and not statistical sampling. We used the non-parametric Mann-Kendall test to test for increasing monotonic upward or downward trends and the non-parametric Theil-Sen slope estimator to assess the upward/downward rate in annual tree cover loss in each ecoregion (Sen 1968, Kendall andGibbons 1990). Compared to ordinary least squares regression estimators the nonparametric Theil-Sen estimator of slope is less sensitive to outliers (Sen 1968, Theil 1992. Furthermore, to indicate the spatial variability in temporal trends of tree cover loss across the study area we used the nonparametric Mann-Kendall test to test for upward/ downward monotonic trends in annual time series of tree cover loss for each grid cell corresponding to our climate dataset (≈4 km). We then classified trends as significantly upward, significantly downward or no significant trend based on the Kendall's tau statistic. Positive Kendall's tau statistic values indicate upward trends, whilst negative values indicate downward trends. Trends were significant at p-values ⩽ 0.10.

Patterns of tree cover loss across ecoregions during periods of notable drought events
To characterize interannual climate variability, we generated time series graphs as well as maps of standardized climate anomalies developed from the gridMET climate dataset. Climate anomalies are effective indicators of regional meteorological fluctuations, including droughts, by removing the underlying spatial variability in long-term climate data and emphasizing the temporal deviations (Crockett and Westerling 2018). To examine the spatial patterns of tree cover loss, we computed pixel-level anomalies for each grid cell for each year of the study period. We summarized tree cover loss pixels into annual area (km 2 ) time series for each grid cell. We calculated the mean annual loss for the same reference period of our climate data, 1986-2015, for comparisons. Anomalies were calculated as departures from the reference mean and then expressed as percentage area change relative to same reference mean, as well as percentage area change relative to 1985. An anomaly value of zero indicates tree cover loss that year equals the reference mean. More positive anomaly values indicate higher levels of tree cover loss, such that 100% indicates twice the area of the reference period mean tree cover loss area. Negative anomaly values indicate tree cover loss area lower than the reference mean. We also calculated tree cover loss anomalies at the ecoregion scale and for the entire study area following same procedure already described above. Tree cover loss anomalies ⩾100% were considered as tree cover loss hotspots.

Climate variables most associated with fire-induced tree cover loss for each ecoregion
For each ecoregion and the entire state of California, we determined the climate variable (and season) with the strongest association with fire-induced tree cover loss during the study period. Thus, we used Spearman's rank correlation to assess the association between each of the climate anomalies and annual area of fire-induced tree cover loss. Correlations were considered significant at p-values ⩽ 0.05.

Temporal trends of climate variables most associated with fire-induced tree cover loss for each ecoregion
We used the non-parametric Mann-Kendall test to test for presence of monotonic trends in annual time series of the two climate variables most associated with fire-induced tree cover loss for each ecoregion.

Temporal trends in annual area of tree cover loss among ecoregions
We found significant upward trends in annual total tree cover loss, fire-induced tree cover loss, non-fire related tree cover loss and fire-induced tree cover change in condition for the entire state (figure 2 and table 2). The Theil-Sen slope estimator indicated that the rate of change of increase in tree cover loss was 18.3 km 2 year −1 across the state (table 2). All components of tree cover loss showed a sudden rise early in the study period in 1987. Thereafter, tree cover remained relatively low until another spike in 2007-2008; then a brief slowdown in 2009-2012 and then a rapid increase from 2013 to end of the study period (figure 2). Annual area of total tree cover loss was highest in 2018 (1901 km 2 ), second highest in 2015 (1556 km 2 ) and third highest in 2008 (1549 km 2 ). The rate of change of increase in non-fire related tree cover loss was 4.5 km 2 year −1 (Theil-Sen slope estimator). Annual area of non-fire related tree cover loss was highest in 2015 (808 km 2 ), second highest in 2013 (679 km 2 ) and third highest in 1987 (672 km 2 ). The rate of change of increase in fire-induced tree cover loss was 14.1 km 2 year −1 (Theil-Sen slope estimator). Annual area of fire-induced tree cover loss was highest in 2018 (1371 km 2 ), second highest in 2008 (1035 km 2 ) and third highest in 2015 (748 km 2 ). The rate of change of increase in fire-induced tree cover change in condition was 3.0 km 2 year −1 (Theil-Sen slope estimator). Annual area of fire-induced tree cover change in condition was highest in 2008 (436 km 2 ), second highest in 2009 (277 km 2 ) and third highest in 2018 (276 km 2 ).  (table 2). Theil-Sen slope estimator indicated that the rate of change of increase in tree cover loss was highest in Sierra Nevada (4.11 km 2 year −1 ),  Tree cover loss anomalies were calculated as annual departures from the reference period (1986-2015) mean (or 'normal'), and then expressed as a percentage to this normal. Thus, anomaly value of zero (dashed black horizontal line) represents area of tree cover loss that is equal to the normal. Positive anomaly values represent area of tree cover loss higher than the normal, such that 100% indicates twice the size of the normal. Negative anomaly values indicate area of tree cover loss lower than the normal, and (b) tree cover loss metrics as a percentage of initial tree cover area in 1985.
followed by Klamath Mountains (4.03 km 2 year −1 ) and the Cascades (1.66 km 2 year −1 ) (  Figure 5 indicates drier and hotter conditions in California around 2012-2016 that were reflected in unusually high VPD, temperature (both minimum and maximum) and low precipitation. Spatial patterns of VPD analysis indicate that the following years had unusually high atmospheric dryness that were widespread across the state : 1987-1988, 1992, 1994, 1996, 2000-2004, 2007-2009, 2012-2016 and 2018 (figure 6). The years 2012-2016 were particularly associated with much warmer temperatures.

Patterns of tree cover loss across ecoregions during periods of notable drought events
The 1987-1988 drought was concentrated in the northern part of the study area and was spatially coincident with very high levels of tree cover loss anomalies (or hotspots, as areas with tree cover loss anomalies exceeding 100%). These hotspots were mostly in the Klamath Mountains and

Climate variables most associated with fire-induced tree cover loss for each ecoregion
Across the state, patterns of fire-induced tree cover loss, and indeed total tree cover loss, were strongly linked to interannual climate fluctuations with higher losses coincident with protracted droughts that were often associated with abnormally high temperatures, especially in 2007-2009 and 2012-2018 (figure 8 and supplementary material figure S2 g). The hotter drought during 2012-2016 was accompanied by sharp increases in both fire and non-fire related tree cover loss and tree cover change in condition. Across California, the largest annual area of total tree cover loss and fire-induced tree cover were recorded in 2018.   Table 3 indicates the top two climate variables most associated with fire-induced tree cover loss for each ecoregion. VPD had the strongest positive association with fire-induced tree cover loss in the Sierra Nevada and Southern California Mountains (r = 0.72, 0.48, respectively). Additionally, VPD had Figure 7. Time series maps showing the spatial pattern of total tree cover loss anomalies across California for the period containing the greatest amount of tree cover loss, 2008-2019. This period also covers notable dry and warm years during the study period. Anomalies were calculated as departures from the 1986-2015 mean, and then expressed as percentage area change relative to this reference period mean. Anomaly value of zero indicates tree cover loss that equals the reference mean. More positive anomaly values indicate higher levels of tree cover loss, such that 100% indicates twice the area of the reference period's mean tree cover loss area. Negative anomaly values indicate tree cover loss area lower than the reference mean.   figure S2). Across the entire state, fire-induced tree cover loss had the strongest positive association with minimum temperature of the fire season and VPD of the water year (r = 0.79 and 0.75, respectively). From 2008 onwards we observed upward trends in all components of tree cover loss across California that were consistent with rising minimum temperatures and VPD (figures 2, 8 and supplementary material figure S2(g)). All correlations were significant at p-values less than 0.05.

Discussion
Overall, California experienced substantial loss in tree cover between 1986 and 2019, and the rate of the loss accelerated rapidly towards end of the study period. Although both fire and non-fire related tree cover loss had upward trends across the state, the former increased at a much faster pace than the latter. Patterns of tree cover loss were strongly linked to interannual climate fluctuations with higher losses coincident with drought and warmer years. A prolonged warmer drought during 2012-2016 was associated with record tree cover loss. The rise in non-fire related tree cover loss that were observed in the later part of the study period, especially between 2013 and 2015, was mostly the effects of moisture stress compounded by widespread pest outbreaks Bales 2019, Madakumbura et al 2020b) rather than timber harvesting. Analysis of Wang et al (2022) suggests that the widespread reduction in tree cover during that period was not primarily harvest-driven. More particularly, the conifer-dominated Sierra Nevada ecoregion experienced one of the worse tree mortalities during that period. The drought acting in concert with bark beetle outbreaks resulted in elevated tree mortality across southern Sierra Nevada (Pile et al 2019, Keen et al 2020, Reed and Hood 2021. Between 2014 and 2017 almost 49% of trees died within a network of national forests across central and southern Sierra Nevada (Fettig et al 2019). In the Klamath Mountains, tree mortality was very high with the area of canopy decline during that drought reaching three times greater than the long-term mean area of canopy decline (Bost et al 2019).
We found areas of high tree cover loss (or hotspots) that temporally varied across ecoregions. These hotspots were initially concentrated mostly in the northern ecoregions but progressively shifted southwards, especially to Sierra Nevada. Despite the general upward trends of tree cover loss, some locations had downward trends. Even though we did not examine tree cover gain/recovery in this study, we believe that slowdown in timber harvest contributed to the downward trends. Historical annual timber harvest shows drastic decline in timber harvest in later part of the study period across ecoregions (see section 3 in SM for details). Large-scale timber harvesting had traditionally been concentrated in ecoregions such as the Coast Range and the Klamath Mountains but the total volume of harvest greatly diminished during our study period, especially from public land (Raumann and Soulard 2012, Sleeter and Calzia 2012, Sohl 2012, Marcille et al 2020, California State Assembly 2022.
Tree cover loss hotspots were spatially and temporally consistent with severe drought conditions mostly observed in the last decade of the study period. Even though 2018 did not have the harshest drought conditions, it was the year with the largest area of total tree cover loss, including fire-induced tree cover loss, in the entire state and the Klamath Mountains.
Here, we infer that the harsher drought conditions of 2012-2016 may have resulted in additional fuels that later promoted more severe fires leading to high fireinduced tree cover loss in 2018.
Drought-induced tree mortality can develop large, contiguous areas of combustible biomass, which can interact with subsequent fires to create higher potential for more catastrophic fires (Ruthrof et al 2016, Stephens et al 2018. Goodwin et al (2021) described the mechanism by which tree mortality and increasing fuel aridity due to climate change could aggravate the intensity of fires. They demonstrated that in two fires in southern Sierra Nevada and Colorado Rocky Mountains, substantial tree mortality and drier fuels could have doubled potential fire radiative energy. Indeed, much larger severe fires have been reported in the Sierra Nevada in the aftermath of the 2012-2016 drought. In a recent study, Stephens et al (2022) found that high concentration of dead biomass created from the drought and bark beetles together with high live tree densities resulted in large, severe wildfires with adverse, landscape-scale fire effects. Moreover, extensive tree deaths in the 2012-2016 drought may have enhanced productivity of understory grass and forbs in some areas thereby increasing fine fuel continuity capable of catalyzing more severe fires in 2018. These findings indicate that increased warming and frequent droughts accelerated tree cover loss across the state during the study period. Further, these results portend that the compounding effects of frequent climatic extremes and changing disturbance regimes may magnify the probability of critical transitions in ecosystems across the study area.
The protracted drought conditions during 2012-2016 was characterized by hotter temperatures, higher VPD, and lower precipitation. This prolonged warmer drought triggered extreme moisture stress in trees leading to substantial tree deaths (Asner et al 2016, Bost et al 2019, Goulden and Bales 2019, Taylor et al 2019, Madakumbura et al 2020a. We found increased tree cover loss in response to drought conditions across several ecoregions. VPD had a strong positive association with annual area of fire-induced tree cover loss for majority of the studied ecoregions. This result is in line with the finding that rising VPD has contributed to increased burned area in the western United States. (Higuera and Abatzoglou 2021). Using VPD as a proxy for aridity, Juang et al (2022) found that annual forest area burned in the western United States showed exponential response to aridity. VPD was a key driver of high-severity fires in several recent large-scale wildfires in California (Safford et al 2022). Our temporal trend analysis indicated an upward trend in VPD across ecoregions and the entire state. Moreover, elevated temperatures promote fire activity by enhancing fuel aridity as well as rapid ignition and fire spread. Higher minimum temperatures signify warmer nights; although nights are usually a time window when fires lose energy and slow down, warmer nights are increasing fire spread . Therefore, the positive association between minimum temperature and fire-induced tree cover loss observed across the entire state, Central California Foothills & Coastal Mountains, Klamath Mountains and Sierra Nevada ecoregions was understandable. Thus, the strong positive relationships between fire-related tree cover loss and both VPD and temperature, as well as the upward trend of these climate variables toward more arid and warmer conditions indicate that climate change may remain a major driver of tree cover depletion in the future. Nonetheless, it is worth noting that due to aggressive fire suppression and exclusion, forests were previously overstocked with trees  leading to water stress (Goulden and Bales 2019, Cui et al 2022. Tree death from water stress and fires remove many of the competing trees and potentially allow for increased water availability for remaining trees in the future. Arguably, large trees present in forests that have died back because of drought may be more resilient to future droughts.

Conclusion
Drought and wildfire associated tree cover loss mostly surged in California during the last decade of the study period; a trend consistent with several other wildfire studies in the western United States (Parks and Abatzoglou 2020, Dwomoh et al 2021, Iglesias et al 2022. Our findings indicate that frequent hotter and drier climate in the future may trigger more rapid tree cover loss, catalyzed by more severe fires, with the potential of engendering ecosystem regime shifts across ecoregions. Notably, fire-unrelated tree cover loss also increased during the 2012-2016 drought period, especially in the Cascades, Eastern Cascades Slopes & Foothills, Klamath Mountains and Sierra Nevada. Aside from escalating fire activity, warmer droughts promote insect-driven tree mortality which may in turn foster larger and more severe fires. In this study, we did not attempt to break out the tree cover to grass/shrub change resulting from timber cutting, which was a major activity in the northern ecoregions, and a bit in the Sierra Nevada (Raumann and Soulard 2012), in the early part of the of the study period. Although we found substantial tree cover loss during drought years, especially following the multi-year warmer droughts of 2012-2016, we notice that response to drought varied by ecoregion, contingent on ecoregion-specific historical climate variability, drought tolerance and disturbance history. The differences in response of tree cover across the studied ecoregions indicate that ecoregionspecific mitigation and adaptation strategies would be beneficial. Such strategies may benefit from consideration of historical disturbances, the ecoregion's sensitivity to disturbance types, and potential climatevegetation-disturbance feedback loops.

Data availability statement
All datasets used in this study are publicly accessible in the sources indicated in the text.
The data that support the findings of this study are openly available at the following URL/DOI: https:// doi.org/10.5066/P9W1TO6E.