The three study sites were Niwot Ridge (US-NR1), USA , Saskatchewan (CA-Obs), Canada and Sodankylä (FI-Sod), Finland . All of these sites are evergreen coniferous forests. The Canadian and Finnish sites are in the boreal zone, and Niwot Ridge is a subalpine forest. Further details about the sites are given in Table . All sites have eddy covariance flux observations as well as a tower-mounted in-situ SIF instrument. The Sodankylä site is part of the ICOS network (https://www.icos-cp.eu/, last access: 1 September 2025) and the North American sites are part of AmeriFlux (https://ameriflux.lbl.gov/, last access: 1 September 2025).

Although all of these sites exhibit a strong seasonal cycle in vegetation activity, this varies due to differences in latitude and elevation. The forest in CA-Obs exhibit a strong seasonal cycle, characterized by low levels of photosynthetic activity between October and March. Although US-NR1 is located at a lower latitude than the other study sites, high elevation conditions result in a pronounced seasonal cycle, including winter below freezing temperatures. Photosynthesis in the forest at US-NR1 is active from May to September, with the shoulder season to winter occurring in October and November, and the spring recovery occurring in April and May. FI-Sod is located 100 km north of the Arctic Circle. Therefore, the winter radiation drops to zero, and air temperatures are low. Spring recovery occurs in April and May. The photosynthetically active period is from June to August and photosynthesis ceases in September and October.

At the North American sites, SIF was observed with PhotoSpec in two different spectral regions. The red region of PhotoSpec is between 680 and 686 nm, and the far-red region is between 745 and 758 nm. These observations were made with a 2D scanning telescope. The retrieval method was based on the Fraunhofer line method . The field of view (FOV) was 0.7°. At US-NR1 a typical measurement involved scanning from nadir to the horizon in steps of 0.7° at two different azimuth directions . At CA-Obs three vertical scans at three different directions (35° W, 0° N, 35° E) were done in sequence . The PhotoSpec retrievals were filtered using a Normalized Difference Vegetation Index (NDVI) based threshold to ensure that only vegetation observations were used. In this study we averaged over all the three scans. Further details of these observations can be found in (US-NR1) and (CA-Obs). At US-NR1, observations were available from June 2017 to June 2018. At CA-Obs we used observations from the whole of 2019 and 2020.

In Sodankylä, the observations were made using a FloX box (JB Hyperspectral Devices, Düsseldorf, Germany) (https://www.jb-hyperspectral.com/products/flox/, last access: 1 September 2025). These observations were used to retrieve the SIF in the O2B band at 687 nm in the red region and theO2A band at 760 nm in the far-red region. The retrieval method used to process the data was the improved Fraunhofer line method . In our study, we used observations close to nadir from June 2021 until the end of the year. The FOV is 25°. The different wavelengths of the retrieved SIF signals by the instruments are shown in Fig. , along with the SIF spectrum from observations of Scots pine needles in Hyytiälä, southern Finland .

Net ecosystem exchange of CO2 was measured by the eddy covariance method. At CA-Obs the measurement height was 25 m, and the anemometer was CSAT3 (Campbell Scientific Inc., Logan, UT, USA) and the gas analyzer LiCor LI-7200 (LiCor Inc., Lincoln, NE, USA) . In spring 2019 the eddy covariance observations at CA-Obs were out of commission and until mid-2019 the GPP data was based on gap–filling. At US-NR1 measurements were made at 21.5 m height with the CSAT3 and an LiCor LI-6262 gas analyzer . At FI-Sod, measurements were made at 25 m with a Gill HS-50 sonic anemometer (Gill Instruments, Lymington, UK) and a LiCor LI-7200 gas analyzer . Flux partitioning and gap–filling at CA-Obs was done as described in and at US-NR1 using the method with the R package REddyProc . Gap–filling and partitioning of the measured net ecosystem exchange flux to gross primary production (GPP) and total ecosystem respiration was done at FI-Sod following . In Sodankylä, the fraction of absorbed photosynthetically active radiation (fAPAR) was measured using PQS1 instruments (Kipp & Zonen; Netherlands) . Four of these sensors were installed below the canopy. These observations, together with aboveground canopy observations, were used to calculate fAPAR.

The TROPOMI is aboard the Copernicus Sentinel-5P mission and has been providing data since 2018 . TROPOMI provides global, continuous spatial sampling with a daily revisit time because it has a nearly sun-synchronous orbit with a 16 d repeat cycle and a 2600 km swath . The pixel size at nadir was 3.5×7.5 km² at the beginning of the mission, decreasing to 3.5×5.5 km² after August 2019 . We used the TROPOSIF product derived from the 743–758 nm window, at 740 nm . The retrieval methodology was based on the Fraunhofer line in-filling principle and a data-driven method was used .

In this study we used a sampling area of 0.5° × 0.5° around the two study sites, CA-Obs and FI-Sod. This corresponds to an area of approximately 56 km × 33 km at CA-Obs and 56 km × 22 km at FI-Sod. According to the 2019 MODIS MCD12C1 data set , the land cover for CA-Obs in this region was 65 % woody savannah and 26 % evergreen needleleaf forest, with minor contributions from mixed forests and croplands. For FI-Sod, the land cover was 83 % woody savannah and 17 % savannah. We also tested a smaller sampling area extending 0.25° around the site. In this smaller region, the land cover around CA-Obs was 45 % evergreen forest and 53 % woody savannah, with a small contribution from mixed forests. For the FI-Sod site the smaller region comprised of 92 % woody savannah and 8 % of savannah. We did not use TROPOMI data for US-NR1 because TROPOMI observations only covered part of the in situ observational period. TROPOMI's Level 2 cloud fraction product was applied for a strict cloud filtering, removing all SIF data for which the cloud fraction exceeded 0.2, as recommended by . For our analysis, we used daily averages that we had calculated from instantaneous values.

The QUantifying Interactions between terrestrial Nutrient CYcles and the climate system (QUINCY) model is a terrestrial biosphere model that can be run on a single site or on larger scales, such as regional or global. QUINCY uses plant functional types (PFTs) to describe different ecosystems. In site level simulations, each site is represented by a single PFT. Canopy can have up to ten layers. A brief description of the model is provided here; further details can be found in .

The complete version of QUINCY features fully coupled carbon, energy, nitrogen and phosphorus cycles. The model has a modular structure that allows only some parts of the model to be run. We used the canopy module, which calculates the model's fast biophysical processes, including stomatal conductance, photosynthesis and radiative transfer within the canopy. Influence of the soil is also considered, so that water uptake is constrained by soil moisture given a prescribed, PFT-specific root profile. The leaf area index (LAI) and leaf nitrogen content are prescribed with a constant value in the canopy module (otherwise these would be calculated prognostically inside the model). Leaf stoichiometry, i.e. the nitrogen to carbon ratio, is also fixed in the canopy module. The calculation of leaf chlorophyll from leaf nitrogen is described in Sect. S1.1 in the Supplement.

Photosynthesis is calculated according to . This approach is based on the biochemical model of , but instead of the regular implementation of having the minimum of the two branches limiting photosynthesis (light-limited rate of photosynthesis and carboxylation capacity limited rate), the amount of light-saturated region in the leaf is taken into account. In the non-light-saturated part, photosynthesis is calculated using the light-limited rate of photosynthesis based on the maximum electron transport rate parameter Jmax,25 (the parameter has been scaled to 25 °C). For the light-saturated part, photosynthesis is calculated as the minimum of electron transport rate-limited photosynthesis and the carboxylation capacity limited photosynthesis (determined by the maximum carboxylation capacity parameter Vc(max),25). Photosynthesis is calculated separately for sunlit and shaded leaves in each canopy layer and coupled to the stomatal conductance , described in Sect. S1.2. Evergreen trees in cold environments adapt their photosynthesis during the shoulder seasons as described in Sect. S1.3.

The leaf chlorophyll fluorescence yield was calculated using the model developed by . The equations can be found in Sect. S1.4. We did not change any of the model's default parameter values and these parameters were kept constant across all sites. The only change to the standard implementation of the model was caused by the different Farquhar et al. model formulation adapted for use with QUINCY, which did not require any additional parameters.

The escape fraction describes how much of the emitted SIF signal reaches the top of the canopy. The total canopy SIF can be expressed using the escape fraction fesc as : 2SIF=PAR⋅fAPAR⋅ΦFt⋅fesc where SIF is the observed SIF, PAR is the photosynthetically active radiation, fAPAR is the fraction of absorbed PAR and ΦFt is the chlorophyll fluorescence yield. This section introduces the three different approaches to calculating radiative transfer of SIF in the canopy that were used in this study. The first approach uses the mSCOPE model, which has been incorporated into the QUINCY model. This hyperspectral approach considers the attenuation of the SIF signal within the leaf. Rather than being implemented in QUINCY, the L2SM model uses QUINCY's output. In our case, it utilizes two spectral regions (visible below 700 nm regions and near infrared for above 700 nm). This approach considers the attenuation of the SIF signal within the leaf. The final approach is based on an empirical relationship and estimates the escape fraction using the fraction of absorbed PAR and leaf reflectance obtained from QUINCY, which considers the visible and near infrared regions in the same way as L2SM. All approaches treat SIF emission as a diffuse flux.

The output of the radiative transfer approaches in Sect. 2.5.2–2.5.3 is in flux units, i.e. µmolm-2s-1. To be able to compare the model output with the observations, which are typically in units of Wm-2s-1nm-1sr-1, we need to convert the units of the model output. This procedure was similar as described in and also in Sect. S1.7. The QUINCY run with mSCOPE provides ChlF directly within the observed units, and is therefore independent of this approach.

Sustained non-photochemical quenching (NPQs) is a process that is relevant to evergreen plants. This is another NPQ mechanism in addition to the reversible NPQ that we introduced in the Sect. S1.4 (Eq. S14). In previous work with the Community Land Model (CLM) model , a parameterization for sustained NPQ based on the state of acclimation was developed. We have previously used the concept of the state of acclimation in the QUINCY photosynthesis model (Eq. S5). Following the earlier work and similarly to the state of acclimation, we obtained for KNs: 4KNs=KNs,max1+ebNPQs(S-TNPQs) where KNs,max, bNPQs (°C⁻¹) and TNPQs (°C) are parameters, set to 8.0, 0.5 °C⁻¹ and 5.0 °C, respectively. The difference between this equation and Eq. (S5) is that it has large values in winter, while Eq. (S5) has large values during the summer. S is obtained from Eq. (S4). To estimate the parameters in the Eq. () we used SIF observations from US-NR1 and adjusted the parameter values to achieve the best possible match with our model over the entire observational period. A seasonal cycle of KNs and S at Sodankylä in 2021 are shown in Fig. S5.

When both reversible and sustained NPQ were taken into account, KN was then a sum of them, as 5KN=KNrev+KNs.

All the sites in the study were classified as a boreal coniferous evergreen forests PFT in QUINCY. For this study we took advantage of the modular structure of QUINCY and only used the canopy model in our simulations. The LAI and leaf nitrogen content were set to ensure that the average summertime GPP level corresponded with the observed values. The meteorological data (air temperature, precipitation, atmospheric pressure, vapor pressure deficit, wind speed, short and longwave radiation) required to run the model were obtained from the site measurements. In addition, we used atmospheric CO2 concentration and N deposition, although the N cycle was not active in the canopy model simulations. The canopy model does not require any spinup, and the simulations were performed for the years of the observations. The leaf single scattering albedo used to calculate reflectance for conifers was 0.15 in the visible wavelength region and 0.73 in the near infrared . The soil albedo for all the sites was estimated to be 0.15 in the visible and 0.30 in the near infrared. The LAI values were set to lower than the observed values to better match the observed magnitude of GPP. The LAI was set to 2.5 m2m-2 in CA-Obs, 3.6 m2m-2 in US-NR1 and 1.2 m2m-2 in FI-Sod. When presenting the results in Sect. , we had used the sustained NPQ presentation for the CA-Obs and US-NR1 sites, but not for the FI-Sod site.

The metrics to assess the model performance were coefficient of correlation (r2) , bias and root mean square error (RMSE) as well as its systematic and random components . These formulas have been shown in Sect. S1.6.

For many of the figures we used averaging window of 15 d to smooth out the daily values, so that the seasonal cycle would be easier to distinguish. We showed for the subdaily time scale separately metrics for the model performance on morning (06:00 to 09:30 a.m., local winter time), midday (10:00 a.m. to 01:30 p.m.) and afternoon (02:00 to 05:30 p.m.). All of these times were local winter time.

As the mSCOPE was run inside QUINCY, while the two other approaches were run outside, the wall time calculation differed. For the mSCOPE version, the wall time was calculated as the time it took for QUINCY to simulate one year at FI-Sod. For the L2SM and the LZ approaches, QUINCY was first run for one year at FI-Sod. Then the calculation performed outside QUINCY in Python was added to this wall time value.

SIF modelled without sustained NPQ showed a strong relationship with the absorbed PAR (aPAR) at CA-Obs (Fig. a). A CA-Obs simulation was done to evaluate the performance of the parameterization carried out at US-NR1. QUINCY successfully simulated the aPAR at CA-Obs at midday (Fig. a, Table ). The seasonal cycle was strong, with winter aPAR values around 200 µmolphotonsm-2s-1. The increase towards summer aPAR values started earlier than the observed GPP increase, as low temperatures prevented the spring recovery of vegetation. The increase in aPAR towards summer values started in the first part of the year, much earlier than the increase in SIF values (Fig. ). The simulated SIF, without the described NPQs, followed the behaviour of the aPAR. The simulated SIF with the NPQs was more similar to the observed seasonal behaviour (Fig. b, Table ). While the magnitudes differed between the simulations and observations for SIF, the general timing was better for the simulation with NPQs. The NPQs had a strong influence on the chlorophyll fluorescence yield (ΦF) in the model (Fig. c).

As US-NR1 is located further south than CA-Obs, the aPAR did not exhibit a pronounced seasonal cycle (Fig. S10a). The simulated SIF without NPQs followed the seasonal cycle of aPAR. To simulate the seasonal variation observed in SIF, it was necessary to include NPQs in the modelling. However, the formulation used for NPQs delayed the spring recovery in 2018 by too much. The formulated NPQs also slightly affected the summertime values, which is unlikely to happen physiologically in reality. Including the NPQs improved the simulation results considerably in terms of r2, RMSE and bias (Table 4).

FI-Sod is located north of the Arctic Circle, where the aPAR exhibits a pronounced seasonal cycle (Fig. S11a). The aPAR caused the simulated SIF values to begin increasing in February, which was well before the start of the active vegetation period (Fig. S11). The aPAR observed by the above- and belowground canopy PAR sensors showed much better resemblance to simulations than the aPAR estimated from the FloX measurements (Table ). The temperature response of the chlorophyll fluorescence yield (ΦF) showed further the pronounced difference depending on whether the NPQs was included or not (Fig. S12). There were no differences between the sites. The temperature response of the ratio of daily far-red SIF to GPP showed that this ratio increased in cold temperatures at the North American sites, as observed and simulated (Fig. S13). The simulated values showed much larger SIF : GPP ratios than the observed values in low temperatures. The range of temperatures covered by the observations at FI-Sod was not wide enough to show such variation.

Noticeable differences were found comparing the relationship between GPP and far-red SIF in the observations and simulations with L2SM for June and July (Fig. ). The observations presented equally high far-red SIF values for CA-Obs and US-NR1, although the observed GPP values were higher at CA-Obs (Fig. a). Both observed GPP and far-red SIF values were lower at FI-Sod compared to the North American sites (Fig. a). The simulated values showed much less scatter for the SIF-GPP relationship than the observations (Fig. b). The highest far-red SIF values were obtained at CA-Obs, although the highest GPP values were obtained at US-NR1. When examining the GPP and far-red SIF light responses (Fig. e, g), it was observed that the simulated far-red SIF values were closely related to the PAR values, whereas the simulated GPP values exhibited greater variability.

We performed a hyperbolic fit on these relationships, determining the parameters a and b of the function y=ax/(b+x) . The fitted lines are shown in Fig.  and the fitted parameter values with their associated uncertainties are shown in Table S5. The goodness of the hyperbolic fits (Table S5) was better for the simulated GPP vs. SIF relationship than for the observed relationship (averaged over three sites r2=0.73 for the simulated and r2=0.45 for the observed.) Also, the RMSE of the fit was smaller for the simulations than for the observations at all three sites. The worst fit behaviour (in terms of r2) occurred at FI-Sod for the observations, which could reflect the fact that the GPP vs. SIF relationship was fairly linear at that site.

Due to the poorer model performance for SIF at US-NR1, the diurnal cycle during the summer was examined in more detail for all three sites. First, we calculated the r2 and RMSE values for the instantaneous values over five days versus the averaged diurnal cycle values over five days. The improvement in the r2 and RMSE values when moving from instantaneous to averaged values was considerable (Table ).

The r2 values of the averaged diurnal cycle were comparable for GPP and SIF in the far-red region at CA-Obs (Table ). On day of year (DOY) 187, the simulated GPP showed an almost sinusoidal diurnal behaviour, resulting from simulation under high irradiance conditions (Fig. a). The observed GPP showed more variation than the simulated GPP during the day. The light response of the observed GPP was much more scattered than that of the simulations (Fig. e). The simulated far-red SIF values with all the approaches were able to capture variations quite successfully on DOY 189 (Fig. c), when variations in radiation occurred.

At US-NR1 the model's performance was significantly lower for GPP than at the other two sites (Table ). This remained the case even after calculating the average over five days. The light response curves of the observed GPP and SIF in the far-red region were quite scattered at this site (Fig. S14e, f). The model performed very well for GPP at Sodankylä (Table , Fig. S15). However, the modelled diurnal cycle of far-red SIF did not capture the observed variation (Fig. S8c, d).

The magnitude of the simulated SIF was larger than the tower-based observations at the two sites (Figs. , S6, S8). A comparison with satellite observations at CA-Obs revealed better agreement between the simulated and observed magnitudes (Fig. c) than with proximal sensing (Fig. b). In 2019, the seasonal cycle of TROPOSIF was smoother than those of the simulated SIF and PhotoSpec observations at the site. In 2020, the seasonal cycle was smoother in the PhotoSpec observations, with the simulated seasonal cycle becoming more consistent with TROPOSIF. The simulation results and the TROPOSIF are shown on different scales because the magnitude of the winter TROPOSIF observations is below zero due to the retrieval method. However, the minimum of the simulated SIF is zero, so using different scales helps to illustrate the seasonality of the two observations together. The simulations for July-August overestimated the observations by 34 %.

At FI-Sod (Fig. S16) the TROPOSIF time series expanded also to spring, thus covering also for the period not available from the FloX observations. There was considerable variation in the TROPOSIF observations during spring (shaded region in Fig. S16, from 9 April to 17 May), and by the time the values increased for the last time in May (the vertical line in Fig. S16 showing 29 May), the observed GPP was already at higher level than in early spring. The variation in TROPOSIF and the simulated SIF followed the same pattern during the summer. Overestimation of the simulations compared to TROPOSIF observations for the July–August period was 41 %.

The performance metrics for the simulated SIF against the TROPOSIF were better at CA-Obs than at FI-Sod (Table S6). This was partly due to the fact that the sustained non-photochemical quenching was not simulated at FI-Sod. The FloX observations alone were not sufficient to assess that, as the spring was missing from them. However, as can be seen from the Fig. S16a, c, the increase in TROPOSIF should have started earlier than it did here. Despite the large differences observed in the springtime at FI-Sod, the RMSE and bias were quite similar at the two sites for the larger region (a 0.50° × 0.50° area surrounding the region). When constrained to a smaller region, the RMSE and bias worsened for both sites, but the r2 increased at FI-Sod, albeit still quite small.

All of the approaches that we tested considerably overestimated the in situ observed SIF (Table S4). The leaf level model was consistent across all our approaches. In that model, we used only the default parameters, which were held constant across all sites. The leaf level model provided the chlorophyll fluorescence yield, which was applied differently by each radiative transfer approach to obtain the top of canopy value.

Therefore, the reason for the overestimation of the SIF could originate from either the leaf level model or the radiative transfer calculation. The leaf level model provided the Ft, which was consistent with values in the literature (e.g., the original model formulation, ; another modelling study for US-NR1 ; and site level observations ). However, there are other types of estimate for these values, including for some of these sites. For example, provided a higher leaf level chlorophyll fluorescence yield value from MoniPAM observations. However, the magnitude is subject to the choices made during the post-processing of the data, so it is difficult to compare the simulated value with these observations.

It is likely that the overestimation originated in the radiative transfer component of the model. Despite testing different approaches, they all resulted in an overestimation of SIF. Two approaches that included attenuation of the SIF signal inside the leaf (mSCOPE and L2SM) resulted in smaller overestimation than the approach that did not account for attenuation (LZ). Additionally, the LZ approach used here assumed black, or non-reflective, soil. Some recent developments may help to overcome this issue . The simulated far-red region SIF showed lower overestimation than the simulated red region SIF. This model behaviour suggests attenuation issues in the red region. Modelling SIF is also related to modelling of absorbed PAR. At US-NR1, the simulations overestimated absorbed PAR; at the other two sites, however, the bias in aPAR was smaller (Table ). It has been noted, that parameters related to aPAR are important for SIF modelling .

Overestimation of SIF has also occurred in other modelling studies of evergreen conifer forests , and our results are close to the model average shown in a model comparison study conducted at US-NR1 . Preliminary model tests with QUINCY on other ecosystems did not reveal such significant discrepancies between the magnitude of the simulation results and in situ observations (results not shown). Therefore, it is possible that the characteristics of the in situ sampling in this type of ecosystem result in smaller SIF signals than expected. A comparison with satellite observations did not reveal such a large overestimation of the simulated SIF. It should be noted that other processes may also play a role, such as reversible NPQ. The leaf level model that we used has been parameterized for cotton . used leaf level MoniPAM observations from Hyytiälä to parameterize reversible NPQ, thereby improving the performance of their SIF model. This could be a way to improve modelling for new ecosystems.

First, we tested different ways of describing the radiative transfer of the SIF to determine a robust method for feasible calculations in a large-scale model. When considering the r2 metric for the different approaches, we found that the simple LZ method did not perform significantly worse than the more sophisticated approaches (Tables , S2, S3). This justifies the rather simple approaches previously used in the radiative transfer of SIF e.g.,, but it contrasts with some other studies . The mSCOPE model performed similarly to the other approaches at all sites and often provided the most accurate estimates. However, its longer computation time (20 times longer than the other approaches in our comparison) renders it impractical for large scale applications . The radiative transfer model in mSCOPE is based on SAIL, which was originally developed for croplands . As the radiative transfer model in QUINCY was also originally developed for croplands, neither model was designed to consider the unique characteristics of radiative transfer in the conifer forests.

The structure of the vegetation affects the observed SIF signal . This is something that could be addressed by including vegetation NIRv, a near-infrared reflectance, in the analysis. NIRv interacts with the canopy in a manner that is very similar to that of the far-red region SIF. Including NIRv in the analysis to help interpret the SIF signal would enable the attribution of structural effects observed in the SIF signal . This analysis could be conducted using in situ or spaceborne observations.

The single leaf scattering albedo of forest plant functional types in QUINCY is based on a study by . The values commonly used to calculate reflectance and transmittance in terrestrial biosphere models have been criticized, and some new approaches based on more extensive data have been proposed . Since reflectance was also used in the calculation of the LZ approach calculation, it would also influence these results. The assumption of equal reflectance and transmittance in QUINCY is debatable and could be further developed, as discussed in . Using only two radiative bands (visible and near infrared) may introduce biases in the results at certain wavelengths.

The L2SM and the LZ approaches require SIF spectra for unit conversion from modelled to observed units (see Sect. ). For the LZ approach, we used spectra measured in a Finnish Scots pine forest. To extend this approach to other PFTs, different measured SIF spectra would need to be used. This approach is limited in terms of generalizability because these spectra differ between species . Therefore, using a single spectrum for a PFT may introduce uncertainties. Furthermore, the spectral shape of SIF emission changes under stress conditions. For example, photosystems I and II respond differently to stress , which will further limit our approach, and require careful investigation of the stress effects. For the L2SM approach we used a theoretical estimate of the in vivo spectrum. Further testing with observed in vivo spectra would also improve this approach. The simulated clumping index at the FI-Sod (Fig. S1) was close to the observed estimates . The seasonal cycle was consistent with the observed increase in the clumping index with increasing solar zenith angle .

Using the Fraunhofer line method with PhotoSpec instruments makes the measurement less susceptible to atmospheric attenuation than using the oxygen lines with FloX. At FI-Sod, the distance from the soil to the FloX instrument was around 19–20 m, and the distance to the canopy was shorter. Therefore, the measurement distance was less than 20 m, which is the threshold at which the data must be corrected for atmospheric effects . FI-Sod has a sparser canopy than the other sites, meaning the footprint of the observing optical fiber is susceptible to environmental influences other than the canopy, such as the understory. The nearly linear relationship between GPP and SIF in the observations may indicate understory contribution (Fig. a).

The measurements that we used have several sources of uncertainty. The tower SIF observations are relatively new observations and are subject to uncertainties relating to instrumentation, the retrieval method, and the spatial matching of the optical and flux footprints . Our results showed that averaging could be useful for the model evaluation. The eddy covariance observations are subject to uncertainty due to the measuring equipment, the heterogeneity of the footprint, and the stochastic nature of turbulence . Gap-filling introduces further uncertainties to the data .

Comparing the satellite observations to a site level observations introduces several uncertainties. The fact that the TROPOSIF estimates remain close to zero by the end of May at FI-Sod, despite the GPP levels advancing towards summer levels, indicates that these observations must be treated with caution. Clouds make interpreting the signal more challenging, which is why we applied the strict cloud filtering criterion recommended for this type of study. There is a large mismatch in scale between the satellite and flux tower observations. However, constraining the region did not improve the modelling results. This may be because having more data points helps to smooth out the random errors in the satellite observations. When only the pixels where the site was located were selected from the satellite retrievals, there were so few data points that no significant seasonal cycle could be detected (data not shown). Data aggregation for TROPOSIF involves averaging over several viewing geometries because the viewing angle varies due to the wide swath. This introduces additional uncertainty to the data. Using the “daily corrected” values of TROPOSIF would help limit the influence of the uneven number of observations on each day. The land cover classes within the TROPOSIF product that are derived from the MODIS product have issues in the high latitudes, which is a known issue . However, the land cover class does not affect the retrieved SIF because it is only used to identify water bodies and glaciers. Overall uncertainty in the TROPOSIF product has been found to be ∼ 0.50 Wm-2s-1nm-1sr-1 , which emphasizes that the low values during the shoulder seasons for these northern sites are highly uncertain.

Our model only simulates one plant functional type per site, and it is also horizontally homogeneous. Therefore, the influence of the understory is disregarded, even though it may be relevant. A study by found that including clumping in their model significantly improved SIF modeling in CLM compared to simulations without clumping. Failing to consider clumping can lead to significant errors in SIF modelling . QUINCY describes clumping in a relatively simple way. There is potentially room for improvement in SIF modelling if clumping were also included in the radiative transfer of SIF, e.g. in L2SM. In coniferous forests, the challenge is that the clumping of the canopy exposes more ground vegetation visible to optical measurements, affecting the remotely sensed signal .

Compared to the simulated GPP, the observed GPP showed more variation on a subdaily scale at the sites (Figs. , S14, S15). This variation may be due to canopy shading, understorey vegetation and turbulence conditions. The lower performance at CA-Obs in the morning for the simulated SIF may be due to the sun-view geometry and the 3D structure of the canopy, which casts shadows within the spectroradiometer's measurement footprint. These shadows were not reproduced by 1D radiative transfer models. At FI-Sod shadows could also contribute to the discrepancy between FloX observations and simulations (Fig. S15c, d). Designing an observation that excludes the shadow effects would be very challenging at such high latitudes, where the days are very long in the summer. More rigorous modelling tools, such as a 3D description of forest structure, would be required to account for these effects. In reality, directional effects also influence the observed optical signal, but our approach did not account for them . A comparison of simulated and observation based estimates of aPAR from the FloX and PAR sensors revealed that the model's inability to capture certain dynamics observed by the FloX box may be due to its small footprint.

The scatter of the observed SIF against the observed GPP values, as well as the fact that model was unable to fully capture this behaviour (Fig. ), raises questions about the accuracy of the simulated SIF provides reasonable results in water-stressed conditions (an example of dry period shown at FI-Sod in ). However, due to the scattered nature of the SIF observations, having a more comprehensive dataset than that used in this study would be helpful.

Our results showed a strong influence of sustained NPQ at these sites, where the plants cannot use the energy of the incoming radiation due to temperature constraints and winter dormancy (see also ). In general, simulating NPQ has been a challenge to the modelling community, as it is comprises of many processes , and active measurements have been required to quantify it. However, recent advances in the use of spectral imaging of xanthophyll cycle pigments potentially enable quantifying NPQ also from the optical observations , thus making remote sensing of NPQ feasible. Some studies have also combined vegetation indices to assist with the parameterization of NPQ . Some new parameterizations for NPQ based on site-level observations have also become available .

The current results showed that the formulation that was successful at other sites likely caused a too early increase in simulated SIF in spring at FI-Sod, when compared with satellite observations (Fig. S16). The colder temperature and light regime at high latitudes may cause the FI-Sod to exhibit different dormancy levels and coping mechanisms for winter conditions compared to sites in North America. A study of a South Korean evergreen coniferous forest showed that the observed chlorophyll fluorescence yield (ΦF) was lower at low temperatures than our parameterization allowed (Fig. S12). Therefore, it is likely that the parameterization could be improved for the FI-Sod, while still maintaining realistic values for ΦF. Although a similar parameterization for GPP based on the state of acclimation was successful across sites, the same could not be said for a parameterization based on similar principles for sustained NPQ. This may be due to the closer link between ChlF and changes in the pigment pool than between GPP and the pigment pool . The cold protection mechanisms of conifers are complex and include additionally changes in the absorption cross-section of the antenna complexes, as well as downregulation of photosystem II activity, accompanied by the transfer of energy from photosystem II to photosystem I . Previous studies have demonstrated that the ratio of GPP to SIF fluctuates at low temperatures . It is important to model the dynamics of this behaviour in order to understand the coupling between GPP and SIF.

A recent study combining PAM observations worldwide evaluated the photosynthetic capacity of photosystem II . The study showed that this capacity depends more on the temperature regime of the environment where the vegetation grows than on the species of plant. The study also quantified this dependence. The ultimate goal is to obtain a sufficiently general parameterization for the entire coniferous evergreen forest region, using the results of . Using the same model for photosynthesis and ChlF could potentially circumvent this problem. However, the current photosynthesis parameters in QUINCY are directly influenced by nitrogen content and leaf chlorophyll content is also closely coupled to photosynthesis . Leaf chlorophyll content could be useful as a metric related to nitrogen cycling in model evaluation . This is a much-needed metric . Additionally, the amount of leaf chlorophyll influences how much of the SIF emitted from leaves that is attenuated in the canopy within the visible spectrum. Therefore, including both leaf chlorophyll and SIF in a model such as QUINCY would be beneficial for understanding Earth system processes and their temporal variations. The amount of chlorophyll in leaves also affects the shape of the SIF spectrum. As we used a fixed SIF spectrum to convert from the total SIF flux to SIF at a given wavelength , this topic could benefit from some further investigation.