Wildfire‐induced increases in photosynthesis in boreal forest ecosystems of North America

Observations of the annual cycle of atmospheric CO2 in high northern latitudes provide evidence for an increase in terrestrial metabolism in Arctic tundra and boreal forest ecosystems. However, the mechanisms driving these changes are not yet fully understood. One proposed hypothesis is that ecological change from disturbance, such as wildfire, could increase the magnitude and change the phase of net ecosystem exchange through shifts in plant community composition. Yet, little quantitative work has evaluated this potential mechanism at a regional scale. Here we investigate how fire disturbance influences landscape‐level patterns of photosynthesis across western boreal North America. We use Alaska and Canadian large fire databases to identify the perimeters of wildfires, a Landsat‐derived land cover time series to characterize plant functional types (PFTs), and solar‐induced fluorescence (SIF) from the Orbiting Carbon Observatory‐2 (OCO‐2) as a proxy for photosynthesis. We analyze these datasets to characterize post‐fire changes in plant succession and photosynthetic activity using a space‐for‐time approach. We find that increases in herbaceous and sparse vegetation, shrub, and deciduous broadleaf forest PFTs during mid‐succession yield enhancements in SIF by 8–40% during June and July for 2‐ to 59‐year stands relative to pre‐fire controls. From the analysis of post‐fire land cover changes within individual ecoregions and modeling, we identify two mechanisms by which fires contribute to long‐term trends in SIF. First, increases in annual burning are shifting the stand age distribution, leading to increases in the abundance of shrubs and deciduous broadleaf forests that have considerably higher SIF during early‐ and mid‐summer. Second, fire appears to facilitate a long‐term shift from evergreen conifer to broadleaf deciduous forest in the Boreal Plain ecoregion. These findings suggest that increasing fire can contribute substantially to positive trends in seasonal CO2 exchange without a close coupling to long‐term increases in carbon storage.


| INTRODUC TI ON
Arctic tundra and boreal forest ecosystems are important carbon sinks (Kurz et al., 2013;Watts et al., 2023) that partially offset the build-up of atmospheric CO 2 driven by fossil fuel emissions (Friedlingstein et al., 2022;Liu, Deng, et al., 2022;Ruehr et al., 2023).Several lines of evidence indicate that in addition to the annual mean sink, the seasonality of carbon exchange in northern terrestrial ecosystems, particularly in boreal forests, is increasing in response to multiple global change drivers.For instance, surface measurements from high northern latitude greenhouse gas monitoring stations have shown an increasing amplitude of the annual cycle of CO 2 (Forkel et al., 2016;Graven et al., 2013;Keeling et al., 1996;Lin et al., 2020).Aircraft observations over a 50-year period also demonstrate a 57 ± 7% enhancement of the amplitude in the middle troposphere of high northern latitudes (Graven et al., 2013).Model and remote sensing-based studies suggest that the growing CO 2 amplitude results from enhancements in the annual cycle of net ecosystem production, defined as the difference between net primary production and heterotrophic respiration (D'Arrigo et al., 1987;Forkel et al., 2016;He et al., 2022;Lin et al., 2020;Piao et al., 2018;Randerson et al., 1997;Welp et al., 2016;Yin et al., 2018).Proposed mechanisms for the increasing seasonality of net ecosystem production have focused primarily on photosynthesis responses to rising levels of atmospheric CO 2 (Bastos et al., 2019;Bonan, 1992;Chen et al., 2022), agricultural expansion and intensification (Gray et al., 2014;Zeng et al., 2014), and climate change (Forkel et al., 2016;Liu et al., 2020;Myneni et al., 1997;Piao et al., 2008;Randerson et al., 1999).However, recent findings suggest that existing ecosystem models fail to fully capture the magnitude of observed amplitude trends, emphasizing the need to incorporate longer-term adjustments in species composition and other forms of ecological change within the models (Forkel et al., 2016;Graven et al., 2013;Keeling & Graven, 2021).
One driver that has received less attention is the influence of disturbance on the magnitude and phasing of ecosystem fluxes.Zimov et al. (1999) describe how ecological disturbance can enhance the annual cycle of CO 2 exchange in Siberian Arctic ecosystems and hypothesize that increasing levels of disturbance may contribute to observed increases in the atmospheric CO 2 amplitude through changes in succession and thus, the landscape abundance of different plant functional types (PFTs).Chamber measurements from Zimov et al. (1999) show that disturbed sites exhibit enhanced CO 2 exchange as a consequence of deciduous shrubs and grasses replacing evergreen dwarf shrubs and mosses.The recently disturbed sites have higher rates of mid-summer photosynthesis and an increase in respiration during spring and fall shoulder seasons.At the ecosystem scale, eddy covariance measurements in interior Alaska demonstrate considerable enhancement of mid-summer CO 2 uptake in a 15-year-old aspen (Populus tremuloides Michx) stand as compared to an 80-year-old black spruce (Picea mariana (Mill.)B.S.P.) stand (Welp et al., 2006(Welp et al., , 2007)).Other eddy covariance studies in other boreal forest ecosystems provide widespread support for higher levels of mid-summer photosynthesis in deciduous broadleaf forests in comparison to nearby evergreen conifer forests (Amiro, 2001;Barr et al., 2009;Buchmann & Schulze, 1999;Coursolle et al., 2012;Goulden et al., 2011;Gower et al., 1997;Krasnova et al., 2022;Röser et al., 2002;Ueyama et al., 2019).
In many boreal regions, wildfire plays a key role in shaping the regional distribution of PFTs, which in turn can feedback on wildfire frequency, intensity, and severity (Chapin et al., 2011;Johnstone et al., 2016;Rogers et al., 2015).Black spruce is a dominant species in North American boreal forests that thrives in cold and wet environments with stand-replacing fire (Van Cleve & Viereck, 1981).
However, field observations have shown a loss of resilience in black spruce (Baltzer et al., 2021) and satellite observations provide evidence that the spatial extent of black spruce and other evergreen conifer forest cover in western boreal North America has declined over the past several decades (Wang et al., 2020).
Loss of black spruce and evergreen conifer species is likely driven by multiple mechanisms.Increasing levels of annual burned area (Gillett et al., 2004;Wang et al., 2021), from increases in lightning (Chen et al., 2021;Veraverbeke et al., 2017) and more extreme fire weather (Jones et al., 2022), are altering the age distribution of forests, creating a landscape with younger stands in which black spruce remains in the understory.At the same time, fire severity may be increasing in some areas as a consequence of soil warming and drying, enabling higher levels of combustion of moss and organic soils (Turetsky et al., 2011).Severely burned areas promote the establishment of deciduous broadleaf trees due to favorable soil moisture and nutrient conditions (Johnstone et al., 2020;Mack et al., 2021).
Climate change, including soil warming, also may hinder post-fire establishment and growth of black spruce (Baltzer et al., 2021).
Boreal forest species in an open-air southern boreal forest experiment show larger reductions in growth and survivability under manipulated warming conditions compared to temperate forest species (Reich et al., 2022).Modeling studies suggest future increases in wildfire may favor the expansion of deciduous broadleaf forests in Alaska and other northern boreal regions (Foster et al., 2019;Foster, Shuman, et al., 2022;Mekonnen et al., 2019).
Here we examine the impact of changing wildfire regimes on landscape-level photosynthesis or gross primary production (GPP) in boreal North America to test the disturbance-CO 2 amplitude hypothesis put forth by Zimov et al. (1999).Our focus is on western boreal North America due to the availability of a high-quality land cover time series that was developed from Landsat imagery (Wang et al., 2019) as a part of NASA's Arctic-Boreal Vulnerability Experiment (ABoVE; Goetz et al., 2011).We used solar-induced fluorescence (SIF) from the Orbiting Carbon Observatory-2 (OCO-2; Sun et al., 2018) as a proxy for GPP, because SIF is more sensitive to evergreen conifer productivity than other vegetation indices (Magney et al., 2019;Pierrat et al., 2022;Walther et al., 2016).The history of wildfire perimeters is identified using large fire databases from Alaska (Alaska Interagency Coordination Center, 2022) and Canada (Canadian Forest Service, 2022; Canadian National Fire Database, 2022).Our analysis quantifies post-fire trajectories of PFT composition and SIF for the region as a whole and for individual ecoregions, addressing the question: How does fire influence PFT composition and SIF seasonal dynamics?We assess whether postfire SIF trajectories can be solely explained by the way fire changes the composition of PFTs in recovering stands or if other mechanisms may be important.In a final step, we model the effect of long-term trends in regional vegetation cover, primarily driven by increases in wildfire disturbance, on the seasonal pattern of SIF across the entire domain.

| Study domain and datasets
We conducted our analysis on high-latitude ecosystems within the ABoVE core study domain (Loboda et al., 2019), encompassing an area of 4.09 × 10 6 km 2 in Alaska and western Canada.To focus on areas where boreal forests and wildfires were common, we narrowed our analysis to six EPA Level II ecoregions (Commission for Environmental Cooperation, 1997) within the ABoVE core domain.
These ecoregions were the Alaska Boreal Interior, Taiga Cordillera, Taiga Plain, Taiga Shield, Boreal Cordillera, and Boreal Plain (Figure 1a).To facilitate reporting of burned area, land cover, and SIF dynamics, we combined the six individual ecoregions into a study region with an area of 2.64 × 10 6 km 2 .
We identified burned areas (Figure 1b For the combined region, we computed an annual burn area time series in two steps.First, we removed large water bodies from the fire perimeter polygons from the Global Self-consistent, Hierarchical, High-resolution Geography Database Version 2.3.7 (Wessel & Smith, 1996).Second, all fire perimeters in a given year were aggregated to form an annual polygon of burned area for our study region.The total area of these perimeters provided the estimates of total annual burned area in our domain.
To assess regional trends and post-fire changes in PFT composition, we used the 30 m spatial resolution ABoVE annual land cover product covering the period 1984-2014 (Wang et al., 2019(Wang et al., , 2020;; Figure 1c; Table S1).This dataset was derived by time series analysis of Landsat data and mapped 15 land cover classes.For post-fire succession interpretation, we aggregated the original 15 land cover classes into six combined classes (Table 1): evergreen conifer forest, deciduous broadleaf forest, shrubs, herbaceous and sparse vegetation, wetlands, and barren.Additionally, the water class was used to screen out SIF pixels near rivers and lakes.
The land cover dataset was calibrated and validated using visual analysis of high-resolution imagery (i.e.1-4 m) in 1299 sites distributed across the domain (Table S1), and the annual land cover maps had an overall accuracy of 84.1%.Land cover was classified based on the dominant canopy vegetation.For example, the definition of evergreen conifer forest in this dataset is areas that are dominated by evergreen conifer tree cover (>50%), but this can include areas with up to half of the remaining area being covered by other land cover types, such as shrubs or grass; for more details, please see Table 1 of Wang et al. (2020).
Other variables provided in the level 2 dataset include surface albedo and elevation.We used glint and nadir measurements to maximize the number of soundings available for our analysis, focusing on the March 1 to September 30 period when solar zenith angles permitted the most robust observations.Soundings were filtered using the best and good quality flags provided in the v10r dataset.We did not throw out negative values of SIF as that would have created a bias in our estimates of mean SIF.We followed a procedure to address the presence of negative SIF values following guidelines in Doughty et al. (2022).Specifically, we used the uncertainty estimates (σ) provided with the retrieval to exclude SIF values where SIF + 3σ was less than zero.We excluded SIF soundings that contained more than 10% water area, as identified using the 2013 land cover map described above.Within our combined region, there were 2,963,002 soundings from 2014 to 2021, of which 390,773 were completely within fire perimeters from 1950 to 2021.

| Analysis of fire impacts on multi-decadal trends in land cover in the ABoVE domain
We analyzed net land cover changes for the combined region, within, and outside of 1950-2014 fire perimeters to understand how fires are contributing to changes in PFT composition.We created three sets of time series of each land cover class: one for the combined region, another for the area within all of the fire perimeters from 1950 to 2014, and a third for the area outside of those fire perimeters.
We further analyzed long-term trends in PFT composition within the individual ecoregions.Our analysis builds upon the study by Wang et al. (2020) that identified fire as a major driver of evergreen forest loss.

TA B L E 1
Changes in aggregated land cover classes for the regional study area, within fire perimeters and outside of the fire perimeters.

Land cover classes
Total study region

| Land cover trajectories following wildfire disturbance
To understand how post-fire succession influences PFT composition within our study region and across different ecoregions, we quantified PFT composition using the land cover classification described above.We compiled 30 m land cover pixel statistics for the following stand age (i.e., time since last fire) classes: 2-5, 6-9, 10-19, 20-29, 30-39, 40-49, and 50-59 years.We focused our analysis on areas that burned only once in the large fire databases and excluded fire perimeters that had a size smaller than 1 ha.Prefire controls for PFT composition were established by aggregating pixels in the perimeter of fires that occurred between 1986 and 2021, using land cover observations from two or more years prior to the fire.

| SIF trajectories following wildfire disturbance and for individual PFTs
We created post-fire SIF trajectories to observe changes in the annual SIF cycle as a function of time since fire.Stand age was assigned to each OCO-2 SIF sounding by determining the most recent fire it intersected and calculating the difference between the year of the SIF sounding and the burn year.We only considered soundings where 100% of the sounding pixel area was within a single fire perimeter.Soundings were aggregated by month and for the same sets of stand age classes that we used for the land cover analysis described above.While soundings are available from February through October across the combined region, we analyzed the months of March through September as these months had the most highquality observations and more favorable solar zenith angles.We created pre-fire controls for the annual cycle of SIF by aggregating soundings together in the perimeter of fires that occurred between 2015 and 2021, using only SIF soundings that occurred one or more years prior to the fire event.
We also calculated the annual cycle of SIF for individual PFTs to develop endmembers for our modeling analysis described below.
For this, we aggregated SIF soundings by month where at least 67% of the pixel area was comprised of the primary PFT of interest.We used the land cover year for 2013 as our reference year for this assignment since OCO-2 SIF measurements become available in September of 2014.We used SIF soundings that did not intersect any fire or were within pre-2013 fire perimeters.A total of 2,027,996 SIF soundings were available in the combined region to create these PFT annual cycles.
In a second step, we convolved the forest stand age probability distribution function for a given fire return interval with the SIF cycles we derived from the observations for different stand ages (Section 2.4).To assign a SIF value to a specific forest stand age, we fit a spline through the post-fire observations and assumed that SIF changes were linear until they returned to the pre-fire level in year 100 (Figure S2).The 100-year time scale for the ecosystem to return to a pre-fire state is in line with the conceptual model for black spruce in Van Cleve and Viereck (1981) and analysis of satellite land cover observed by Rogers et al. (2013).In a final step, we computed the relative percent change in the monthly SIF using the annual cycle of SIF estimated from a baseline fire return interval.Our baseline corresponded to a mean level of annual burning observed from 1950 to 1969.A more detailed mathematical description of the model is provided in the SI.

| Fire contributions to long-term trends in PFT composition
To understand how fires have changed the land cover during the Landsat era, we estimated long-term net changes in PFT composition for the study region, within and outside of fire perimeters (Figure 2, Table 1).From 1984From -1988From to 2010From -2014, the area of evergreen conifer forests experienced a net decrease of 11.2 Mha or 10.3% (Figure 2a).Most of the loss occurred within 1950-2014 fire perimeters, where evergreen conifer forest cover decreased by 11.7 Mha (Figure 2b).However, this was partially offset by 0.5 Mha of evergreen conifer forest gain in areas outside of the aggregated set of fire perimeters (Figure 2c).Losses of evergreen conifer forest cover occurred consistently across all the individual ecoregions we analyzed and were strongly associated with areas experiencing fire (Figure S3).Outside of fire perimeters, evergreen conifer forest area remained nearly stable or slowly increased in all ecoregions except for the Boreal Plain, where cumulative losses were 0.7 Mha (6.7%).

| Post-fire trajectories of PFT composition
To understand how PFTs change during post-fire succession, we analyzed land cover as a function of post-fire stand age (Figure 3).
In the first decade after fire, losses of evergreen conifer forests and deciduous broadleaf forests were balanced by an expansion of herbaceous and sparse vegetation, shrubs, and wetlands.Herbaceous and sparse vegetation cover quadrupled immediately after fire, increasing from 13% in the pre-fire control to 52% in 2-to 5-year stands.Shrub cover increased into the following decade, reaching a maximum in 10-to 19-year stands (21%).Deciduous broadleaf forests exhibited a delayed response compared to shrubs, expanding in cover area considerably between 20-to 29-year (7%) and 50-to 59year stands (20%).Evergreen conifer forests displayed the slowest recovery, reaching a maximum in 50-to 59-year stands (48%) that corresponded to about two-thirds of pre-fire levels (68%).The observed pattern of expanding evergreen conifer forest cover is consistent with conceptual models of succession, whereby evergreen conifer cover slowly expands to pre-fire levels as a consequence of a series of light, nutrient, and hydrological cycle feedbacks that allow slower-growing evergreen conifers to outcompete faster-growing shrub and deciduous forest species (Chapin et al., 2004;Johnstone, Chapin, et al., 2010;Mack et al., 2021;Van Cleve et al., 1983).
Successional patterns varied considerably across the different ecoregions (Figure 4).A prominent deciduous broadleaf forest phase during mid-succession was visible in Alaska Boreal Interior and Taiga Plain ecoregions, and was nearly absent in Taiga Cordillera Taiga Shield, and Boreal Cordillera ecoregions.In the southern Boreal Plain, there was little evidence that evergreen conifer forests returned to pre-fire levels following fire disturbance.Instead, fire appears to have initiated a more permanent expansion of deciduous broadleaf forests, with evergreen forest cover reaching a plateau in stands older than 20 years.

| Post-fire successional trajectories of SIF and albedo
SIF increased considerably in 2-to 59-year post-fire stand age classes during mid-summer (June and July) relative to pre-fire controls (Figure 5).The greatest SIF enhancements occurred for 20-to 29year stands, with a 35% increase in June and 40% during July.In older stands, mid-summer SIF declines, but remains elevated relative to pre-fire levels.In 50-to 59-year stands, for example, the SIF enhancement is 27% in June and 30% in July.Similar (but noisier) patterns of post-fire enhancement of mid-summer SIF were observed at the scale of individual ecoregions (Figure S4).To further assess the net impact of fire on SIF across our study region, we created two separate SIF time series for areas within and outside of 1950-2020 fire perimeters.This analysis revealed consistent SIF increases in fire-affected areas during June (26% ± 12%; mean ± 1 standard deviation) and July (14% ± 4%) in the OCO-2 record, and evidence for considerable interannual variability in GPP within our study domain (Figure 6).
Concurrent near-infrared albedo measurements from OCO-2 (right-hand side axis in Figure 5) provided additional evidence for loss of the evergreen conifer overstory in the first few decades following fire.Specifically, albedo during March and April was considerably elevated in early successional stands relative to pre-fire controls.This pattern can be explained by greater snow exposure that accompanies replacement of an evergreen conifer forest canopy with shortstature deciduous vegetation and highlights the challenges of using surface reflectance-derived vegetation indices during snow-covered periods in spring in northern ecosystems (Delbart et al., 2005(Delbart et al., , 2006;;Huemmrich et al., 2021).These observations are consistent with previous analysis of satellite and tower broadband shortwave albedo observations that indicate fires can induce a negative topof-atmosphere radiative forcing during spring, offsetting to varying degrees warming effects from fire-emitted greenhouse gases (Randerson et al., 2006;Lyons et al., 2008;Jin et al., 2012;Z. Liu et al., 2018Z. Liu et al., , 2019;;Potter et al., 2020).

| SIF as a function of PFT
To understand how vegetation type contributes to the overall postfire SIF trajectory, we isolated the annual cycle of SIF for individual PFTs.Deciduous broadleaf forests had the highest SIF during midsummer (Figure 7).The phase of the SIF annual cycle in deciduous broadleaf forests was also forward-shifted relative to other PFTs, with a maximum in June occurring 1 month prior to the peak in the evergreen conifer forests.Shrub and wetland vegetation classes also exhibited SIF values that exceeded evergreen conifer forest levels during June, July, and August.However, during spring these PFTs, along with herbaceous and sparse vegetation all had SIF levels that were below those observed in evergreen conifer forests.The lower spring SIF in these PFTs was likely a consequence of snow that partially buried low-stature vegetation and delays in leaf expansion in areas dominated by deciduous shrubs (e.g., Salix L.).Although noisier, elevated levels of mid-summer SIF in deciduous broadleaf forests and shrubs relative to evergreen conifer forests were visible in most ecoregion groupings (Figure S5).

| Modeling post-fire SIF using land cover observations
We modeled the trajectory of post-fire SIF using linear combinations of PFT composition as a function of stand age (Figure 3) and PFTspecific annual cycles of SIF derived from the OCO-2 observations (Figure 7).For most stand age classes, our model (Figure 8a) yields higher R 2 and lower root mean square error (RMSE) values compared to a null model where PFT composition was invariant (Table S2); however, there are several notable discrepancies.First, the model predicts that SIF should peak in the older stands, 50-59 years.This prediction contrasts with the observed maximum of mid-summer SIF that occurs in 20-to 29-year stands and the subsequent decline in older stands.Second, the model overestimates SIF for pre-fire stands during mid-summer.These model-data mismatches suggest that some, but not all, of the post-fire SIF enhancement can be attributed to changes in PFT composition.

| Increasing wildfire as a driver of trends in SIF
We performed additional modeling work to understand how a changing fire return interval in the region would alter the distribution of stand ages and thus the annual cycle of SIF.As a back-ofthe-envelope estimate, we note that the annual burn area has nearly  9c).

| Long-term changes in PFT composition as a driver of trends in SIF
We also predicted how the long-term changes in PFT composition (Figure 2) will influence the annual cycle of SIF using the PFT endmembers shown in Figure 7.We estimated that in the combined region from 1984 to 2013 when high-quality land cover data were available, SIF declined during the spring shoulder season (by 5.7% in March, 3.6% in April, and 0.3% in May), increased during summer (by 1.0% June, 1.2% in July, and 0.9% in August), and declined during fall (by 0.1% in September; Figure 10).
The declining SIF signal in spring from the analysis of the land cover trends is consistent with the direction and magnitude of changes predicted from the stand age distribution model shown in Figure 9.For June and July, predictions of SIF change inferred directly from the land cover data are smaller than what our stand age distribution model predicts.These differences may be a consequence, in part, of the stand age distribution model assuming that forest ages reach a steady state for each fire return interval.In the observed record, this is not the case, with climate change continuing to intensify and modify fire regimes (e.g., Zheng et al., 2023).The lower estimates from the land cover-driven model may also reflect a series of post-fire changes in plant physiology and edaphic properties that are not captured solely using PFT endmembers derived from domain-wide averages.

| Fire as a driver of changing species composition
Examining the patterns of post-fire succession across the different ecoregions (Figure 4), several different distinct patterns emerge.
In Alaska Boreal Interior and Taiga Plain ecoregions, deciduous forests become an important component of the landscape in 40to 59-year stands.This pattern is consistent with successional models in which evergreen conifers represent a mature ecosystem state (Johnstone, Chapin, et al., 2010;Van Cleve & Viereck, 1981) and faster-growing aspen, willow (Salix L.), and birch (Betula The monthly time series of OCO-2 solar-induced fluorescence (SIF) in western boreal North America.The orange line shows the average of all SIF retrievals within the set of 1950-2020 fire perimeters.The dark green line shows the average of all SIF retrievals in unburned areas.The two domains that are plotted are complementary and sum to the total study area domain shown in Figure 1.
L.) trees establish in early and mid-successional stages on welldrained soils that can support higher water and nutrient demands (Mack et al., 2021).In other ecoregions, including Taiga Cordillera, Taiga Shield, and Boreal Cordillera ecoregions, deciduous broadleaf forests are nearly absent in intermediate-aged stands.In these areas, the successional pattern is broadly consistent with the poorly drained lowland black spruce chronosequence model described by Van Cleve and Viereck (1981).In this model, evergreen conifer recovers directly without passing through a deciduous forest phase, in some cases because low fire severity leads to thicker residual duff layers, that in turn, more easily desiccate aspen seedlings (Johnstone, Hollingsworth, et al., 2010).A third pattern of post-fire change in PFT composition occurs in the Boreal Plain ecoregion, which is the southernmost ecoregion in our study area.In this ecoregion, fire appears to initiate a longterm shift in plant composition, with broadleaf deciduous forests replacing evergreen conifer forests extensively in 30-to 59-year stands.More data are needed in the form of longer satellite time series to ascertain the robustness (and permanence) of this apparent ecosystem transition.An idealized representation of the three fire-initiated pathways and their regional expression is shown in Figure 11.
Our analysis suggests there are at least two mechanisms by which fire is contributing to a long-term decline in the cover of The annual solar-induced fluorescence (SIF) for individual plant functional types (PFTs) in western boreal North America.All valid SIF retrievals that had an interior composition of a given PFT exceeding 67% within the retrieval footprint were averaged together to create these PFT-specific endmembers.
evergreen conifer forests in western boreal North America.The first is the long-term increase in annual burned area (Figure S6) that is pushing more of the landscape into early and mid-successional states where shrubs and deciduous broadleaf trees are more abundant.A second mechanism is connected to a more permanent replacement of evergreen conifers by deciduous broadleaf forests (Figures 4 and 11) that may be a consequence of poor post-fire seedling establishment from warmer (and drier) summers (Baltzer et al., 2021;Frelich et al., 2021;Reich et al., 2022), the compound effect of multiple disturbance agents (Anoszko et al., 2022;Liu, Riley, et al., 2022), or interactions between warming and nutrient availability that favors the growth strategy of deciduous trees (Foster, Shuman, et al., 2022;Mekonnen et al., 2019;Walker et al., 2023).Together, the rapid decline of evergreen conifers in fire-affected areas and the slower recovery (if any) of evergreen conifers in undisturbed areas (Figure 2; Figure S3) provide evidence that the boreal landscape is far from a steady state, likely as a consequence of increasing fire and warming.
In our analysis, we did not consider the influence of other disturbance drivers, such as harvest or insect outbreaks, which are known to have important impacts on ecosystem structure and carbon cycling (Foster, Wang, et al., 2022;Wang et al., 2020).In the southern part of our domain, increased prevalence of drought (Ma et al., 2012;Michaelian et al., 2011;Peng et al., 2011), inabundance of invasive species from temperate tems (Frelich et al., 2021), and changes in the range and abundance of insects (Kurz et al., 2008;Malmström & Raffa, 2000) may have had additional impacts on the loss of evergreen conifers observed in the Boreal Plain ecoregion.

| Fire as a driver of increasing SIF and GPP
One of the challenges of understanding GPP and terrestrial carbon fluxes at a regional scale in boreal forests is the sparsity of the data acquired from field campaigns and eddy covariance towers; however, satellite measurements can help to fill in gaps (Schimel et al., 2015(Schimel et al., , 2019)).In the past decade, SIF emissions have been retrieved from the analysis of high-resolution spectra of Fraunhofer lines found near the O 2 A-band from radiometers on several satellites using the principles of the Fraunhofer line discrimination method (Frankenberg et al., 2011;Guanter et al., 2012;Joiner et al., 2011;Köhler et al., 2018;Mohammed et al., 2019).SIF has also notably been important for boreal ecosystems where greenness in evergreen conifers may become decoupled from GPP as a consequence of physiological processes in long-lived needles that enable cold hardening and drought tolerance (Adams et al., 2004;Magney et al., 2019;Pierrat et al., 2021).As a consequence, SIF has been shown to yield a higher correlation with GPP than other vegetation indices for evergreen conifer forests (Magney et al., 2019;Pierrat et al., 2022;Walther et al., 2016).
Here we show that fire increases SIF, at a regional scale, in part, by increasing the abundance of shrubs and deciduous broadleaf trees in intermediate-aged stands.The observed enhancements were as high as 40% during July across the full study area, and are broadly consistent with enhancements in GPP and the annual cycle of net ecosystem production observed in early-intermediate-aged stands from tower observations (Amiro et al., 2010;Coursolle et al., 2012;Goulden et al., 2011;Ueyama et al., 2019;Welp et al., 2006).The Comparison of a model of post-fire changes in solar-induced fluorescence (SIF) with observations.Panel (a) shows the model of SIF created by combining information on post-fire changes in plant functional type (PFT) composition (Figure 3) with SIF endmembers for individual PFTs (Figure 7).Panel (b) shows the SIF observations to enable a visual comparison (these data are the same as shown in Figure 5).The model captures some, but not all, of the post-fire SIF enhancements visible in the observations.The model-data mismatch suggests that not all post-fire changes in SIF can be attributed solely to changes in PFT composition.Other potential mechanisms that may structure observed SIF enhancements after fire (but are not represented in our simple model) include changes in nutrient availability (Alexander & Mack, 2016) and active layer dynamics (Viereck et al., 2008).Additionally, the model overestimation of SIF for pre-fire stands may be indicative of a higher probability of fire ignition and spread in denser black spruce stands, which comprise only a subset of the areas used to derive the domain-wide evergreen forest SIF endmember. ( | 13 of 23 expanding presence of deciduous vegetation over evergreen vegetation in post-fire ecosystems also likely contributed to declines in SIF during spring and fall shoulder seasons. Increases of SIF in post-fire stands are consistent with past work identifying that places with significant summer greening trends, as measured using normalized difference vegetation index (NDVI), occur within recovering intermediate-aged stands (Fiore et al., 2020;Sulla-Menashe et al., 2018;Wang & Friedl, 2019).Postfire enhancement of SIF is also consistent with past work documenting increases in enhanced vegetation index (EVI; Jin et al., 2012;Lyons et al., 2008).However, use of reflectance-based indices in the boreal regions can be more limiting due to sensitivity to snow and surface water (Delbart et al., 2005(Delbart et al., , 2006;;Huemmrich et al., 2021), whereas SIF is less sensitive to these effects (Frankenberg & Berry, 2018;Pierrat et al., 2021Pierrat et al., , 2022)).Analysis of fire impact on SIF in boreal forest ecosystems of Northern China and Siberia reveals similar post-fire enhancements (Guo et al., 2021), even though fire regimes and PFTs differ considerably from those in western North America (Rogers et al., 2015).
A key objective for future work is to connect the SIF observations reported here to an ecosystem model of carbon fluxes.In this context, understanding the SIF-GPP relationship and how it varies for different boreal PFTs is an important step.Magney et al. (2019) showed that SIF more accurately captures the annual cycle of GPP of an evergreen forest at Niwot Ridge compared to changes in NDVI or absorbed photosynthetically active radiation alone.The linear relationship between SIF and GPP strengthens when the timescale of integration lengthens from hourly to monthly intervals at the site level (Magney et al., 2019;Pierrat et al., 2022).The linearity of the SIF-GPP has been observed from satellites at larger spatial and temporal scales, and some work suggests that a single scale factor (slope) between SIF and GPP may be applicable within and across biomes (Li et al., 2018;Li & Xiao, 2022;Sun et al., 2017).The linearity and the universality of the relationship may arise at regional scales because SIF integrates information related to both structural and physiological processes (Magney et al., 2020;Sun, Wen, et al., 2023).However, ecosystem structure can also influence observed SIF by altering the escape ratio and the incidence of reabsorption of fluorescence, which is known to influence the slope of the SIF-GPP relationship in different biomes (Zeng et al., 2019;Zhang et al., 2020).Through the inte- relationship was relatively consistent for evergreen needleleaf forests, deciduous broadleaf forests, and shrublands.The congruence in slopes for these three PFTs, accounting for reported uncertainties, lends support to the notion that the changes in SIF driven by fire in boreal forests of western North America, as described here, likely translate into proportionate impacts on GPP.

| Implications for understanding changes in the annual cycle of atmospheric CO 2
The strong enhancement of mid-summer SIF in post-fire stands highlights the importance of fire in structuring the composition of PFTs and the annual cycle of GPP in boreal forest ecosystems.From 1984 through 2013, mid-summer SIF within our study area was expected to increase by about 1.2% during mid-summer (July) solely as a consequence of changes in PFT composition from our modeling analysis (Figure 10a).Over this same interval, the Utqiagvik (Barrow) seasonal amplitude increased by about 18% (0.6% year −1 × 30 year; Graven et al., 2013).If we assume similar landscape-level changes in vegetation cover across Eurasia and other areas that contribute to the footprint of seasonal exchange influencing the Utqiagvik station (Kaminski et al., 1996;Lin et al., 2020), and that GPP enhancements are proportional to SIF enhancements, then fire-driven changes in land cover may contribute to about 7% (1.2% / 18%) of the cumulative CO 2 amplitude change.Evidence for a similar increasing trend in fire disturbance in Siberia, for example, comes from satellite analysis of long-term trends in burned area and fire impacts on atmospheric carbon monoxide (Zheng et al., 2023).With these simplistic assumptions, our work suggests that disturbance-driven increases in GPP may be responsible for a small, but non-negligible, component of the CO 2 amplitude change.This provides direct support for further consideration of the Zimov et al. (1999) disturbance mechanism in CO 2 amplitude attribution studies.
Several lines of evidence that disturbance-driven contributions to the amplitude may be even larger.Randerson et al. (1999) showed that increases in net ecosystem carbon uptake in the early part of the growing season were more consistent with changes in the shape of the annual CO 2 cycle in northern high-latitude regions than increases in net ecosystem exchange that were distributed in proportion to the mean annual cycle.Enhanced ecosystem uptake during the early part of the growing season is likely driven by spring warming (Butterfield et al., 2020;Keeling et al., 1996;Myneni et al., 1997;Randerson et al., 1999).Our analysis here reveals a second mechanism that likely reinforces the effect of spring warming.Specifically, post-fire SIF increases are not symmetric with the annual cycle of SIF in pre-fire stands.Increases in SIF in 2-to 59year stands are concentrated during June and July, even though in pre-fire stands SIF is nearly the same during June, July, and August (Figure 5).These observations suggest that fire-induced changes in source is larger prior to leaf out and the transition to a carbon sink is more abrupt (Welp et al., 2006;Zimov et al., 1999).Greater summer carbon inputs and warmer fall temperatures may allow for greater respiration to occur during the fall (Piao et al., 2008;Welp et al., 2016); fire-driven changes in vegetation cover would likely amplify these effects as a consequence of decreases in fall GPP (Figure 5).
Fire-induced changes in soil properties can also drive important changes in the magnitude and timing of GPP and soil respiration.Nutrient availability is known to increase immediately following fire, in part from the concentration of residual combustion products on the forest floor and also from the decomposition of plant roots killed by the fire (Dyrness et al., 1986;Van Cleve et al., 1983).Higher rates of decomposition, mineralization, and nitrification are also expected from higher soil temperatures after fire (Alexander & Mack, 2016;Wan et al., 2001), that in turn, can support higher rates of nutrient uptake by deciduous broadleaf tree species and other fast-growing species found in early stages of post-fire succession (Mekonnen et al., 2019).Although fireinduced soil warming can increase heterotrophic respiration from mineral soil layers, it is important to note that losses of surface organic soil layers can also reduce soil respiration fluxes during the growing season (Amiro et al., 2003;Czimczik et al., 2006).In areas underlain by permafrost, fire-induced soil heating may also increase water availability by extending the active layer depth (Viereck et al., 2008).Together, these edaphic changes are a critical aspect of the biophysical environment that supports high levels of net primary production (Mack et al., 2008)  mid-successional stages.A lack of representation of these soil processes is likely a key reason why our simple PFT-endmember model underestimates post-fire SIF enhancements in Figure 8, and thus likely underestimates long-term trends in Figure 10.
On decadal timescales, there is also a potential synergy between increasing levels of fire disturbance and the response of vegetation to warming and rising levels of CO 2 .With a warming climate, deciduous forests may store more carbon (Wang et al., 2023) through a response of an earlier bud burst (Gunderson et al., 2012).In contrast, the above-ground production of black spruce has been observed to decline with increases in warming (Dusenge et al., 2020;Hanson et al., 2020).Both evergreen conifer forests and deciduous broadleaf forests can benefit from CO 2 enhancement (Norby et al., 2005), although GPP and growth responses for aspen and other deciduous broadleaf species are often stronger than those for black spruce (Hanson et al., 2020).
Thus, fire-driven changes in species composition may increase the sensitivity of GPP to important global change drivers at a landscape scale.
Increasing levels of fire disturbance may be increasing GPP and seasonal CO 2 exchange yet decoupling trends in these fluxes from longer-term changes in carbon storage.Carbon storage within boreal forest ecosystems can be reduced if high-carbon ecosystems are replaced with low-carbon storage ecosystems (Koven, 2013) or if losses from harvest and fire are not balanced by gains in recovering forests (Wang et al., 2021).For example, organic carbon storage declines when thick moss and organic soil layers that are often present in black spruce stands are removed by fire (Potter et al., 2023;Walker et al., 2019).Build-up of organic soil layers is often slower in aspen and birch stands because of more rapid leaf decomposition and pulse of leaf litter that may impede moss growth (Johnstone, Chapin, et al., 2010;Van Cleve & Viereck, 1981).For high-severity fires, however, carbon losses from below-ground pools may be partially offset by more rapid rates of accumulation in above-ground biomass pools due to the faster growth and larger size of many deciduous broadleaf forests (Mack et al., 2021;Madani et al., 2021;Mekonnen et al., 2019).

| Uncertainties and future directions
We used level-2 OCO-2 SIF as a proxy for GPP; however, it is important to recognize that SIF is not GPP (Porcar-Castell et al., 2014), and a thorough exploration of nonlinear SIF dynamics reveals situations where these fluxes will diverge (Maguire et al., 2020;Marrs et al., 2020;Sun, Gu, et al., 2023;Yang et al., 2022).The record is also relatively short for OCO-2 SIF (2014-2021), limiting our ability to get a more robust pre-fire characterization of SIF for different ecoregions.We also note that solar zenith angle and sun-sensor geometry can influence the SIF signal (Doughty et al., 2022;Joiner et al., 2020).Although some of these effects may average out in our study because of the relatively large spatial and temporal scale of our analysis (e.g.Doughty et al., 2022;Magney et al., 2020;Sun, Wen, et al., 2023), more work is needed to understand SIF differences inferred from nadir and glint retrievals (Porcar-Castell et al., 2021).Li et al. (2018) demonstrated that OCO-2 nadir and glint mode SIF-GPP slopes were similar at the daily timescale when aggregating over 64 flux towers.However, the slopes may vary when analyzing individual towers (Zhang et al., 2018).New flux measurements covering a broader range of Arctic tundra and boreal forest land cover types as well as higher temporal and spatial resolution measurements of SIF from space are needed to develop more robust SIF-GPP relationships (e.g., Cheng et al., 2022) and, more broadly, to understand how climate change is modifying ecosystem function in northern biomes (Liu et al., 2020).
There is also a need to extend high spatial resolution land cover time series to better understand land cover change across the full pan-boreal domain of Eurasia and North America.This is critical for improving the interpretation of atmospheric CO 2 annual cycle trends from northern observing stations since these records widely integrate information from the two continents.Further, while the large fire database records are vital for extensive space-for-time substitution chronosequence studies exploring SIF and other land surface properties, they are known to have issues in accuracy going back in time (Foster, Wang, et al., 2022)

| CON CLUS ION
We analyzed Landsat-derived time series of land cover and OCO-2 measurements of SIF to better understand post-fire changes in PFT and GPP in boreal forest ecosystems of western North America.
We find that fire-induced changes in PFT contribute to increases in mid-summer SIF in 2-to 59-year stands.We also observe a forward shift and sharpening of the SIF annual cycle as a consequence of deciduous shrubs and trees replacing evergreen conifer trees.At the landscape scale, our analysis and modeling suggest there are two pathways by which increases in fire in western boreal North America may be increasing GPP.The first is through increasing fire inducing a shift in the stand age distribution of forests, creating a regional mosaic with more early and mid-successional stands.A second path is through a long-term shift in vegetation composition from fireclimate interactions, whereby evergreen conifer forests are more permanently replaced by deciduous broadleaf forests.Together, the land cover and SIF analyses reported here provide support for the idea that disturbance from fire can enhance GPP and, as a result, ) using fire perimeters from the Alaska Large Fire Database (Alaska Interagency Coordination Center, 2022) and the Canadian National Fire Database (Canadian National Fire Database, 2022).The Alaska fire record covered the period 1940-2021, while the Canadian fire record covered 1917-2020.For Canada, the National Burned Area Composite dataset (Canadian Forest Service, 2022) was used to extend the time series through 2021.Post-fire successional trajectories of land cover composition and SIF were developed, using wildfire perimeters from 1950 to 2021, with stand age measured in year since fire.F I G U R E 1 Study domain and visualization of the fire, landcover, and solar-induced fluorescence (SIF) datasets used in our analysis.(a) The six level 2 ecoregions within the Arctic and Boreal Vulnerability Experiment (ABoVE) core study domain that form the basis of our study region are shown in color: Alaska Boreal Interior (blue), Taiga Cordillera (red), Taiga Plain (pink), Taiga Shield (purple), Boreal Cordillera (light blue), and Boreal Plain (green).The total study region boundary is shown in black.(b) 1950-2021 fire perimeters from the Alaska large fire database, the Canadian national fire database, and the national burned area composite.(c) Aggregated vegetation cover classes derived from the Wang et al. (2019) land cover time series are shown for the representative year of 2014.(d) Filtered OCO-2 SIF soundings from the level 2 product are shown from March 1 through September 30 for the 2014-2021 period.For display purposes, all maps are projected to the EPSG:4326 coordinate reference system, and the land cover map is resampled from 30 m to 1 km resolution.Map lines delineate study areas and do not necessarily depict accepted national boundaries.
understand how post-succession changes in PFTs alter the annual SIF cycle, we created a simple model to predict monthly SIF as a function of stand age using only information about how PFT composition changes post-fire (Section 2.3) and the idealized SIF annual cycles for each PFT (Section 2.4).Our model had the form Ax = b, where each row is a different month and each column of the matrix A is the average annual SIF curve for a single PFT and (A has dimensions of 7 months × 6 PFTs).The vector x is the fractional contribution of each PFT present in each stand age class and has dimensions of 6 × 1. x is derived directly from our analysis of land cover.b is the annual cycle of predicted SIF for a given stand age class (b has dimensions of 7 months × 1).We estimated the annual SIF curve for each stand age class using the different vectors of PFT composition for x (Section 2.3).We aimed to quantify how much of post-fire SIF trajectory could be explained solely as a function of landscape-level changes in PFT composition.This analysis did not consider influence of post-fire changes in nutrient or water availability on the expected SIF signal within an individual PFT class.We also combined the annual SIF cycles from the individual PFTs and long-term trends in the land cover time series (Section 2.2) to estimate how land cover change between 1984 and 2013 (30 years) has influenced regional SIF.
In contrast to the observed increases in mid-summer, post-fire SIF declined during spring and fall shoulder seasons.During April, SIF declined by 57% in 2-to 5-year stands and gradually recovered to pre-fire levels in 50-to 59-year stands.Post-fire SIF decreased during May in 2-to 5-year stands (by 45%), although older stands exceeded pre-fire levels after about four decades.Reductions in SIF were also observed during fall months, with declines of 7% during August and 35% during September in 20-to 29-year stands relative to pre-fire controls.Together, the mid-summer enhancement and shoulder-season reductions in SIF led to an overall sharpening and earlier shift in the phasing of the annual cycle in the first six decades after fire.F I G U R E 2 Land cover change in the western boreal North America study region from 1984-2014.Panel (a) shows the plant functional type (PFT) time series in the full study region.Panel (b) shows the PFT time series within 1950-2014 fire perimeters.Panel (c) shows the PFT time series outside of this set of fire perimeters and is estimated using the land area complement to the region shown in panel (b).
of OCO-2 SIF data with flux tower GPP observations, Zhang et al. (2020) quantified SIF-GPP slopes for various PFTs.This analysis revealed discernible differences among tropical forests, C4 grasses, and other vegetation types.However, the slope of the SIF-GPP F I G U R E 9 Modeled annual solar-induced fluorescence (SIF) as a function of fire return interval.Panel (a) shows the distribution of post-fire fire stand ages predicted by the model for a given fire return interval.Panel (b) shows the annual cycle of SIF predicted by the model for a given fire return interval.The curves in panel (b) were generated by combining the stand age information shown in panel (a) with post-fire trajectories of SIF derived from the observations (and shown in Figure 5 and Figure S2).Panel (c) shows the percent change in monthly SIF relative to a fire return interval of 284 years derived from the mean level of burning during 1950-1969.The dashed red lines show the fire return interval corresponding to the mean burning level during 2002-2021 (152 years) and 1950-1969 (284 years).

F
Predicted changes in solar-induced fluorescence (SIF) from long-term trends in plant functional type (PFT) composition.We modeled longterm changes in the SIF time series using the observed trends in PFT composition from 1984 to 2013 shown in Figure 2 and PFT-specific SIF endmembers shown in Figure 7. Percent changes are cumulative and reported relative to the estimated annual SIF in 1984.Each line represents a different month.Percent change in SIF are shown for (a) the entire western boreal North America study region, (b) in areas within fire perimeters in the study region, and (c) in areas outside of fire perimeters.and land cover may contribute to a forward shift in the annual cycle of GPP.Fire-driven replacement of conifers with deciduous shrubs and trees also reduces SIF during the spring and fall shoulder seasons.This additional compression of the annual cycle of GPP may further increase CO 2 amplitude changes by increasing the magnitude of ecosystem carbon losses associated with plant and soil respiration during spring and fall.In boreal evergreen conifer ecosystems, increasing soil and plant respiration losses during spring associated with soil thaw(Goulden et al., 1998) and root flushing(Curiel yuste et al., 2004) are often offset at an ecosystem scale by concurrent gradual increases in moss and overstory canopy GPP.In chamber and eddy covariance observations, this type of gradual transition during spring contrasts with the pattern observed in deciduous broadleaf ecosystems, where the magnitude of the net ecosystem exchange and enhanced seasonal variation in net ecosystem production during early and F I G U R E 11 A conceptual diagram describing different post-fire successional models in North American boreal and their influence on SIF and photosynthesis.The successional model column illustrates expected changes in shrub, deciduous broadleaf forest and evergreen conifer forest cover following fire.The theory/evidence column describes key papers providing evidence for the different conceptual models.The regional expression column represents an attempt to classify ecoregions according to their dominant successional pathway, drawing upon the observations shown in Figures 3 and 4.This figure was adapted from a combination of illustrations shown in Johnstone, Chapin et al. (2010) and Van Cleve and Viereck (1981).
, adding uncertainty when quantifying patterns of land cover and SIF change in older stands.Higher quality Landsat and Sentinel-2 burned area time series (Canadian Forest Service, 2022; Skakun et al., 2022) will be important for reducing uncertainties in future work and extending the analysis into new regions such as Siberia.With a longer and more robust SIF record, in the future it should also be possible to explore post-fire SIF trajectories as a function of different fire severity levels, to better understand how climate change impacts on fire behavior yield decadal-scale legacies in ecosystem composition and function.

The influence of a changing forest age distribution on SIF Increasing
annual burned area within our study region is changing the stand age distribution of forests and decreasing the mean stand age.To understand how changes in forest age structure may influence SIF and GPP, we created an idealized model.First, we Wang et al. (2020) consistent with the results reported byWang et al. (2020)regarding fire-driven forest cover change.