Satellite measurements of SIF are powerful proxies for vegetation growth, and they have been used in many studies to monitor the changes in GPP. Retrieval of SIF from satellite measurements, e.g. from GOME-2A and TROPOMI that offer near-daily global coverage, relies on the relative filling in of solar Fraunhofer lines in the Earth's radiance spectra. The apparent depth of the Fraunhofer lines is not affected by scattering or reflection from clouds or by the Earth's surface. However, if vegetation is present on the surface, the addition of solar-induced fluorescence (SIF) as a spectrally continuous surface light source dampens the relative depth of the lines, and this (relative) filling in of the Fraunhofer lines is a measure of the strength of fluorescence . The spectral-fitting window is made wide enough to cover multiple strong Fraunhofer lines and sufficiently narrow as to limit the presence of atmospheric absorption features. Spectral-basis functions derived from observations over vegetation-free non-fluorescent areas are used to describe the impact of atmospheric interaction and possible instrumental artefacts – the reference set. Accordingly, the SIF signal is computed by fitting the satellite-observed spectrum with a modelled reflectance spectrum assembled from the spectral-basis functions. Uncertainty in the SIF data emerges from measurement noise, the presence of optically thick clouds in the satellite field of view, and the fitting procedure, the latter depending on the quality and representativeness of the spectral-basis functions. To mitigate uncertainty, SIF observations with cloud cover fractions below 0.2 are used.

We evaluate the spatial and temporal trends in TROPOMI SIF over the burned area from February 2018 until November 2020. The monitoring period ends in November 2020 to avoid potential interference from subsequent wildfire events that occurred in December 2020 and affected the studied region . The extent of the eucalyptus fire is identified using the MODIS MCD64A1 Burned Area monthly product (Fig. a), and the impacted burned area is defined as the region encompassing the MODIS burned pixels, indicated by the solid lines in Figs.  and a and b. Next, for comparison, we carefully selected a reference area surrounding the burned area that remained unaffected by the fire. This reference area was chosen to closely match the biogeography of the impacted area, consisting of a mixture of open and closed eucalyptus forest (Fig. b). SIF observations over the reference area are assumed to represent the forest dynamics over the burned area as if the fire had not taken place.

To evaluate temporal trends in SIF over both areas, we categorized TROPOMI level-2 pixels as either “burned” or “reference” based on their spatial overlap with each respective area. Criteria were implemented for each group to ensure pixel exclusivity and minimize interference. Pixels were labelled as burned if the burned-area fraction exceeded 50 %. Pixels were labelled as reference if they comprised less than 30 % of the burned area. Subsequently, within each group, the SIF values were averaged on a monthly basis. Note that throughout the reporting of the results, all uncertainty values are stated in standard error (SE).

Figure b shows the difference in SIF between autumn 2019 (post-fire) and 2018 (pre-fire). Over the area that burned down, autumnal SIF values in 2019 decreased significantly compared those in to 2018. In contrast, SIF at the reference area remained consistent across both years. Furthermore, the lowest SIF values (0.25 mW m-2 sr-1 nm-1) coincide with pixels with burned-area fractions higher than 60 %, whereas SIF is much higher (0.45 mW m-2 sr-1 nm-1) in pixels that had not burned, as illustrated in Fig. c. We conclude that TROPOMI SIF is capable of capturing the strong reduction in vegetation activity over the burned area relative to the unburned reference area, noting that the affected area extends over some 200 km2, equivalent to eight individual TROPOMI pixels.

Figure a shows the seasonal cycle in 2018 (pre-fire) inferred from TROPOMI SIF over the to-be-burned and reference areas. Prior to the fire event, the observed SIF values indicate consistency and strong agreement between the two areas (r=0.97; Fig. S4a). However, shortly after the fire, a substantial reduction in SIF over the burned area becomes evident (Fig. b). This immediate fire-induced impact is highlighted by a striking 72 % decrease in February SIF values compared to those of the reference area. In the subsequent months, the vegetation activity over the burned area remained low but non-zero, with an averaged SIF of 0.26 ± 0.01 mW m-2 sr-1 nm-1 between February and August 2019. During springtime (September–November), vegetation recovery was noticeable, with SIF values reaching 0.43 ± 0.04 mW m-2 sr-1 nm-1 by November. In fact, SIF grew faster in springtime over the burned area than over the reference area, with respective increases of 208 % and 144 %, hinting at accelerated regrowth in a moderately burned area, a trend which has also been reported in Zambia and Alaska .

Next, based on the observed reduction in SIF, we quantify the fire-induced reduction in GPP. Previous studies found space-based SIF observations to be proportional to GPP at the canopy scale. This proportionality is defined by biome-specific linear relationships that are widely used to scale SIF to GPP e.g.. Building upon this recognized relationship between SIF and GPP, we assume that the relative reduction in SIF (Sect. ) represents the loss in GPP (ΔSIF ∝ ΔGPP). Consequently, we express the fire-induced change in GPP over the burned area ΔMB(t) (in gC m-2 d-1) over month t as the product of the relative change in SIF (ΔFB(t)) and the “baseline” GPP (MR (t)) over the reference area (as if no fire had occurred): 1ΔMB(t)=ΔFB(t)MR(t). To compute the SIF-based GPP over the reference area (MR (t)), we first characterize the SIF–GPP relationship using GPP data from a representative eddy covariance site (see Supplement S2). Specifically, GPP data from the nearby Tumbarumba (AU-TUM) site are used to compare against TROPOMI SIF. We obtain the SIF-based GPP over the reference area using the following equation: 2MR(t)=FR(t)MTUMFTUM, where FR (t) represents the SIF evolution over the reference area and MTUMFTUM is the local SIF–GPP relationship over the Tumbarumba flux site. We hold the empirical SIF–GPP relationship over Tumbarumba to be applicable to our study area. This is supported by (i) the similar biogeography of lowland eucalyptus forest in both areas (Fig. b), (ii) the strong spatiotemporal correlation between TROPOMI SIF over Tumbarumba and over the reference area in the year prior to the fire (r=0.92; Fig. S4b), and (iii) the good correlation between FluxSat GPP estimates over Tumbarumba and over the reference area in the year prior to the fire (r=0.79; Fig. S4c). Moreover, the discrepancy between SIF–GPPFluxSat over Tumbarumba and the Nunnett–Timbarra River area is small: on the order of 4 % (Supplement S2.1)

As an alternative, we also compute SIF–GPP relationships over Tumbarumba using TROPOMI SIF and modelled FluxSat GPP. The temporal dynamics of TROPOMI SIF exhibit strong consistency with both the GPP estimated from eddy covariance flux tower data (r=0.91) and the modelled (FluxSat) GPP (r=0.94), providing some confidence in both estimates. However, the SIF–GPP relationship based on the flux tower GPP measurements is 1.6 times stronger than the relationship based on the FluxSat GPP data (Table S2.1, Fig. S5a). This discrepancy likely arises from higher GPP estimates from the eddy covariance flux tower – nearly twice that of the estimates from FluxSat (Fig. S5b) – and inherited uncertainties and biases in the estimates of the FluxSat GPP product. Our findings are consistent with , who also observed magnitude differences up to a factor of 2 between FluxSat GPP and independent FLUXNET 2015 GPP data at some sites. Despite differing footprint sizes between the Tumbarumba tower and FluxSat model, this is not expected to significantly impact the uncertainty, given the extensive land cover around the tower (Supplement S2.1).

Finally, we estimate the reduction in GPP employing SIF–GPP relationships that use (1) the eddy covariance GPP data and (2) FluxSat GPP data over Tumbarumba – hereon referred to as GPPSIF, ec and GPPSIF, FluxSat, respectively. Additionally, we compute the reduction in GPP from the FluxSat GPP data directly over the burned and reference areas, referred to as GPPFluxSat. Figure S6 shows an overview of the three approaches to compute ΔGPP.

Figure a shows the estimated fire-induced reductions in GPP over time. The SIF-based approaches, GPPSIF, ec and GPPSIF, FluxSat, show a sharp reduction in GPP immediately after the fire. This is followed by smaller GPP deficits in winter, when vegetation activity is low. The FluxSat data (GPPFluxSat) also show a direct local impact of the wildfire on GPP. However, the FluxSat data suggest an almost complete recovery in November, in contrast to our SIF-based results.

When integrated over the period from February to November 2019 and across the burned area, we find SIF-based GPP losses ΔGPPSIF, ec and ΔGPPSIF, FluxSat of 198 and 132 GgC, respectively. The SIF-based GPP losses reflect a reduction of 48 % with respect to GPP over the same period in the reference area. Thus, the estimated SIF-based losses are in the range of 130–200 GgC over the initial 8 months post-fire. For comparison, ΔGPPFluxSat indicates a fire-induced loss of 150 GgC over the burned area between February–November compared to the reference area. The total GPPFluxSat over the burned area during this period is 58 % less than over the same period and area in 2018 (pre-fire conditions).

An uncertainty propagation analysis, described in Supplement S2.1, indicates a total uncertainty of approximately 60 % in the estimated SIF-based GPP losses. This total uncertainty is primarily driven by the use of the SIF–GPP relationship at Tumbarumba, where no fire occurred, to derive our GPP-loss estimate via Eq. () over the Nunnett–Timbarra River area that burned down – introducing an estimated uncertainty of 50 %. This uncertainty was determined by comparing SIF and FluxSat GPP relationships under pre-fire and post-fire conditions (see point 2 in Supplement S2.1). We conclude that estimating changes in GPP based on land-cover changes is possible with TROPOMI SIF in combination with empirical SIF–GPP relationships, at least when such land-cover changes occur over a spatial scale of a few hundred km2.

We first examine the 2007–2012 variability in summertime (June–August) GOME-2A SIF, when vegetation productivity is at a maximum. To evaluate the changes in SIF, we used the first Landsat-derived annual China land-cover dataset (CLCD) . These satellite-based data categorize nine major land types such as cropland, forest, and grassland at a spatial resolution of 30 m. Specifically, we extracted forest-coverage data for each year between 2007 and 2012, aggregated over the 0.5° × 0.5° GOME-2A SIF grid. The conversion to forest, from any type of land cover, was determined by quantifying the change in forest cover relative to the 2007 baseline. We then compared changes in summertime SIF between 2007 and 2012 over two contrasting groups of cells: reforested (n=42) and deforested (n=8). We classify cells as “reforested” if they gained more than 2 % of their area in forest coverage (approximately 50 km2) and as “deforested” (n=8) if they lost forest coverage during this period, as shown in Fig. a. Cells in both groups consistently gained or lost forest coverage with every passing year from 2007 to 2012.

The majority of reforested cells showed an increase in average summertime SIF, whereas most deforested cells experienced a decrease (Fig. b). Figure c shows the clear contrast in SIF trends across cells that gained forest coverage compared to those that lost forest coverage. Reforested cells exhibited a significant increase in SIF of 0.17 ± 0.03 mW m-2 sr-1 nm-1, while SIF decreased by 0.09 ± 0.07 mW m-2 sr-1 nm-1 over the deforested cells between 2007 and 2012. The contrasting trends in vegetation activity indicated by GOME-2A SIF over the reforested and deforested cells are confirmed by co-sampled independent data from the near-infrared reflectance of vegetation (NIRv) and FluxSat GPP (Fig. a and b).

Temporal SIF variability across northern China is driven not only by reforestation alone but also by co-occurring changes in local climate factors that influence plant growth, such as water availability, heat, and the presence of photosynthetically active radiation . To examine the co-occurring impact of climate variability and changes in forest cover, we use satellite-based soil moisture and maximum temperature. Soil moisture data were obtained from the Surface Soil Moisture (Active) product and the daily maximum temperature was derived from hourly 2 m temperature from ERA5-Land data . Both were aggregated to the 0.5° GOME-2A SIF grid. The summers from 2007 to 2012 in Shaanxi and surrounding provinces are characterized by soil moisture values (50 %–55 %; Fig. S7b) that are generally good for plant growth, showed a slight increase over time (1 % yr-1), and had relatively stable (average) maximum temperatures (Fig. S7a). The summer of 2010 was somewhat of an exception, with a maximum temperature more than 1 °C below the multi-annual mean, which we interpret as indicating lower radiation levels. Overall, we expect that interannual variability in climate factors favoured forest growth between 2007 and 2012 in both the reforested and deforested areas, in line with previous studies e.g.. Overall conditions for vegetation growth appear more favourable in cells where reforestation occurred, as maximum temperatures were more moderate (23 °C) and soil moisture was higher (53 %) in the cells with reforestation than in cells with deforestation.

We constructed a simple model that expresses the local year-to-year variability in summertime SIF as a linear combination of soil moisture (S; in %), maximum temperature (T; in °C) averaged over spring to summertime (March–August), and the total forest coverage (A; in km2): Eq. (). This model serves to explore the impact of different climatic drivers on observed SIF, representing vegetation activity, alongside the effect of change in vegetation. The selection of this model is motivated by three factors: (i) soil moisture promotes the thriving of vegetation by enhancing the water availability, (ii) increasing temperature stimulates vegetation activity until an optimal temperature is reached and is a proxy for photosynthetic active radiation, and (iii) total photosynthetic carbon uptake increases with increasing vegetation cover. Note that the climate parameters are averaged over the spring and summer period (March–Aug) to account for the concurrent and lagging effects on the vegetation activity over summer (e.g. ).

We use a multivariate-least-squares regression to optimize the combined fit coefficients that best explain the observed SIF over the reforested and deforested cells. Modelled SIF as a function of the year (between 2007 and 2012), Fmod (t), is expressed as 3Fmod(t)=αS(t)+βT(t)+γA(t)+δ, where α, β, γ, and δ represent the coefficients that minimize the discrepancies (ϵ) between Fmod (t) and the observed GOME-2A time series Fobs (t). The predictors of soil moisture, maximum temperature, and forest cover are indicated by S(t), T(t), and A(t), respectively, over the year t. To ensure sufficient statistics, we pooled the time series of SIF (n=6) from all 50 selected cells, resulting in a single cell-ensemble fit. Overall, the modelled variability in SIF matches the measured variability reasonably well (R2=0.57; n=300). Figure a shows the correlation between the observed (Fobs) and modelled (Fmod) SIF over all cells.

However, the model's performance varied by land-use type across the region (Fig. c). Notably, it demonstrated good agreement between observed and modelled values for cells dominated by grassland (r=0.84) or forest cover (r=0.58) but struggled to represent the dynamics in cropland cells (r=0.27). This suggests distinct differences in the underlying dynamics and complexities in converting cropland areas into forest. Previous studies reported that croplands were often located in drier areas with less forest cover to begin with, reducing the vigour with which vegetation increases in response to reforestation .

We therefore exclusively fitted our final model to data from the 36 cells dominated by grassland or forest cover (n=216; Fig. b and c). By excluding cropland cells, the fit significantly improved, evident from the higher R2 of 0.72. The improved correlation between the observed and modelled SIF suggests that the model better captures the underlying dynamics. Note that the impact of different factors on the long-term vegetation dynamics is ecosystem-specific; therefore, our model is unique to the fitted area and period. Next, to evaluate the explanatory power of including forest coverage in our model, we constructed an alternative model that solely fits the first two terms on the right-hand side of Eq. (), i.e. the climate parameters, to the observed summertime SIF. Doing so allows us to compute the adjusted R2 of our model (Eq. ) and the alternate model to enable fair comparison between the models containing three and two predictors. The alternate model had an adjusted R2 of 0.66. With the inclusion of forest-cover data, the model's performance improved, achieving an adjusted R2 of 0.71 – suggesting the added value of including changed forest coverage in conjunction with climate variability in the model.

The model's fit coefficients presented in Table  demonstrate, as hypothesized, that increases in the predictors positively influence SIF. Soil moisture and forest-cover variability explain most of the variability in GOME-2A SIF, as indicated by the scaled values of the fit coefficients (which consider predictor standard deviations). This is also expected, given the relatively strong increases in soil moisture (+5 %) and in forest cover (+97 km2) compared to the modest changes in maximum temperature over the provinces from 2007 to 2012. The effect of changes in soil moisture cannot be disentangled from the effect of changes in forest cover on vegetation activity. However, our results show a positive impact of the forest-cover increase on vegetation activity. This is likely occurring in tandem with the changes in soil moisture because soils in regions with increased tree cover tend to hold on to their soil moisture better. We also tested a model including an additional (fourth) interaction term between soil moisture and reforestation, but this did not lead to an improved R2, adjusted between modelled and observed SIF. We conclude that vegetation activity following reforestation efforts increases most over the temperate areas in northern China classified as grassland and forest, with already relatively high soil moisture. On average, these areas show high covariance between changes in soil moisture and changes in forest cover, indicating that both factors explain the increase in summertime SIF.