The study area covers Finland and ranges from 19° E to 32° E, and from 59° N to 71° N (Figure 1). According to Köppen’s climate classification, the area belongs to the temperate coniferous-mixed forest zone with cold and wet winters. The average temperature of the warmest month exceeds 10 °C and the average temperature of the coldest month is below −2 °C [35]. Finland’s vegetation can be divided into four phytogeographical zones: the hemiboreal, southern, middle and northern boreal zones [36] (Figure 1). The main land cover type is forest (with 44% coverage of evergreen, deciduous and mixed forest), followed by shrubs and herbaceous vegetation, inland waters and wetlands (Supplementary Materials, Table S1). ENF is the dominant forest type, except for areas north of 68° N where DBF is dominating (Figure 1b). The main tree species of forests are Scots pine (Pinus sylvestris) (67%), Norway spruce (Picea abies) (22%), and birch (10%), including silver birch (Betula pendula) and downy birch (Betula pubescens). More than half of the forests are pure stands (55%); stands with some mixing cover 31% of the forested area, and 13% of forest stands are mixed [37].

The net ecosystem CO₂ exchange (NEE) between atmosphere and vegetation can be directly measured by the micrometeorological eddy covariance (EC) method [38]. The 30-min GPP value was here obtained by subtracting the modelled respiration (R) value from the NEE observation (GPP = NEE-R) utilizing standard flux partitioning procedures [39,40]. By the same parameterisation, the NEE and GPP time series were gap-filled for comparison with the model results. In this study, we used EC data from two boreal Scots pine forests, Sodankylä in northern Finland (67°21’N, 26°38’E) [41] and Hyytiälä in southern Finland (61°51’N, 24°17’E) [42] (Figure 1a), from years 2001–2010. Supporting meteorological data from both flux sites were used in the site level JSBACH model simulations (Section 2.4.4). The starting date of the vegetation active season (ENF_SOS_EC) was determined for the two sites from the continuous EC data. ENF_SOS_EC was obtained as the day on which the CO₂ uptake permanently exceeds 15% of the growing season maximum uptake level. The method has been described in more detail in a previous study by Böttcher et al. [33].

We used phenological observations of the bud break of birch (Betula pubescens Ehrh.) from 8 sites for the evaluation of satellite observations of start of season in DBF (DBF_SOS_sat) in Finland. The sites are located in the southern (3), middle (1) and northern boreal zone (4) (Figure 1, Table 1). Bud break observations were obtained from the Finnish Forest Research Institute (now the Natural Resource Institute Finland) for the period from 2001 to 2008. The phenological observations at Aulanko and Saariselkä had already stopped in 2006. According to the observer’s instructions [43], bud break is observed when at least 50% of buds have broken throughout the tree crown. Phenological observations were carried out twice a week at the respective sites. The monitored trees (five) are of local provenance and were preferably naturally regenerated. The microclimate of the site must be representative for the local climate conditions. Monitored trees may have been changed from year to year until 2007. Further details on observations are given in Kubin et al. [43].

We used daily Moderate Resolution Imaging Spectrometer (MODIS) time series for the detection of the start of season in ENF and DBF. The pre-processing of the MODIS time series and the extraction of start of season dates will be described in the following sections.

As described in the previous section, the method for the detection of DBF_SOS_sat was validated earlier in Siberia [70] for boreal forests with a different species composition than in Finland. As described in Section 2.5.2, we made small changes to the original method. Therefore, we used bud break observations of birch from phenological sites (Section 2.3) for evaluation of DBF_SOS_sat in Finland. The DBF_SOS_sat at the coarse grid for the evaluation of JSBACH model results is not appropriate for comparison with ground observation. Thus, the original MODIS NDWI time series with a pixel size of 0.005° by 0.005° were aggregated to 0.05° by 0.05° taking into account pixels with a dominance of deciduous and mixed forest (>50% coverage, according to National CORINE Land Cover) and excluding pixels that were partially (≥10%) covered by water bodies. DBF_SOS_sat was extracted from the nearest grid cell (at 0.05° resolution) to the phenological observation sites. The number of MODIS pixels that were used in the averaging of the NDWI time series is given in Table 1. The difference between the median of elevation within the grid cell and the elevation of the phenological site was less than 50 m (Table 1). For only two sites, Kevo and Koli, the variation in elevation within the pixel was larger than 50 m.

We analysed the correspondence between satellite and ground observations for all site-years, for each site and for the southern, middle and northern boreal zone, separately. Furthermore, we investigated the correspondence of the year-to-year anomalies between DBF_SOS_sat and the bud break observations. The statistical metrics used in the comparison were the coefficient of determination (R²), the root mean square error (RMSE) and the bias. The equations are given in Appendix B (Table B1). We excluded year 2002, because of missing data in the MODIS archive during the spring period. In addition, due to long periods of cloud cover, the DBF_SOS_sat observations were not available for the following site-years: Aulanko 2001, Kolari 2005, Saariselkä 2001 and Kevo 2007 and 2008.

For the comparison of model- and satellite-derived start of season dates, JSBACH results were projected to the same grid as the satellite SOS (0.25°, see Section 2.5.2). We investigated correspondence between the dates for the whole of Finland and for the boreal sub-regions: the southern, middle and northern boreal zone (Figure 1a), and calculated Pearson correlations, RMSE and bias for the period from 2003 to 2010 separately for the two forest types (ENF, DBF). In order to assess correspondence regarding spatial patterns, we used the multi-year mean dates for the whole period for each grid cell. In addition, agreement in temporal patterns was investigated based on the year-to-year anomalies.

For site-level comparisons at Sodankylä and Hyytiälä, the ENF_SOS_mod was extracted from the closest JSBACH grid cell. We calculated Pearson correlations, the RMSE and bias between ENF_SOS_EC and ENF_SOS_mod,loc, and between ENF_SOS_EC and ENF_SOS_mod. We also assessed the agreement between the respective year-to-year anomalies for the period 2001 to 2010.

Comparisons of DBF_SOS_sat with bud break observations including all site-years resulted in an RMSE of one week and an early bias of 0.13 days (R² = 0.75, probability p < 0.001, Number of observations N = 48) (Figure 2a). The site-wise correlation between DBF_SOS_sat and in situ observations were not significant (Table 2), but the number of observations (4–7 per site) was also very small. RMSE and bias ranged between 4 to 10 days and 0 to 5 days, respectively. The smallest RMSE and bias were obtained for northern sites, Saariselkä and Kevo, in Lapland. The largest RMSE (10 days) and bias (5 days early) were both found for Kannus in western Finland. Large deviations (>two weeks) between satellite and in situ observations were observed for Vesijako in 2005 (15 days) and Kannus in 2007 (19 days).

For the southern boreal zone, including sites Aulanko, Vesijako and Koli, the correlation of DBF_SOS_sat with bud break observations was weak (R² = 0.32, p = 0.012, N = 19). The RMSE was 6.4 days and the bias 1.8 days (early). For the four sites in northern Finland: Värriö, Saariselkä, Kolari and Kevo, the correlation was higher than for the sites in the southern boreal zone (R² = 0.62, p < 0.001, N = 22), but the RMSE (6.9 days) and the bias (2.8 days late) were larger. The correlation between interannual anomalies of DBF_SOS_sat and bud break was significant (Figure 2b), but could explain only less than 30% of the interannual variation in the field observations.

We compared ENF_SOS_EC in Sodankylä and Hyytiälä with JSBACH-simulated dates based on the two different sets of meteorological drivers: ENF_SOS_mod,loc and ENF_SOS_mod (Figure 3). ENF_SOS_EC ranged from mid-March to mid-April (33 days) at Hyytiälä, and from the end of April to the beginning of May (14 days) at Sodankylä. The ENF_SOS_mod,loc was too early in comparison with observations (Table 3). The bias was twice as large for Sodankylä as compared to Hyytiälä and correlations between ENF_SOS_mod,loc and ENF_SOS_EC were not significant.

The bias between ENF_SOS_EC and the start of season from the regional run, ENF_SOS_mod, was much smaller than the bias between ENF_SOS_EC and ENF_SOS_mod,loc. Again, the early bias was larger for Sodankylä than for Hyytiälä. Correlations between ENF_SOS_mod-modelled dates with ENF_SOS_EC were significant only for Hyytiälä (Table 3), whereas the year-to-year anomalies correlated significantly for both sites (R² = 0.51, p = 0.001, N = 19).

An investigation of the reasons for differences in ENF_SOS_mod and ENF_SOS_mod,loc in Sodankylä showed that GPP simulated with JSBACH followed closely the development of air temperature during the springs, 2004 and 2008 (Figure 4 and Figure 5). However, the simulated GPP showed a higher increase during warm spells in spring than the GPP from CO₂ flux measurements. This was even more pronounced for the JSBACH GPP simulations with the local driving data than for simulations with regional climate data (Figure 4a,b). The local temperature observations indicated warm spells in 2004 around the day of year (doy) 96, 108 and 119. Therefore, the simulated GPP exceeded the threshold for the start of season earlier than the observed GPP. Consequently, the JSBACH-simulated ENF_SOS_mod,loc and ENF_SOS_mod occurred earlier than ENF_SOS_EC, on doy 96 and 117, respectively. Instead, ENF_SOS_EC was delayed by cold spells that occurred around doy 110 and 115, and it was only observed on doy 121. Modelled snow depth from JSBACH simulations with both local and regional meteorological drivers was lower than the observed snow depth and the snow melted earlier in the model (Figure 4 and Figure 5c).

The difference between JSBACH ENF_SOS_mod and ENF_SOS_mod,loc was smaller in 2008 than in 2004 (only 5 days) (Figure 3 and Figure 5). ENF_SOS_mod occurred on the same day as ENF_SOS_EC observations (doy 123). Again, the local mean daily air temperatures were higher around the time of the start of season than the regional climate model temperatures, thus resulting in a higher GPP increase in JSBACH simulations with local drivers. Similarly to 2004, the snow melted earlier in JSBACH model simulations than indicated by snow depth measurements at Sodankylä (Figure 5c). The simulated GPP seemed to follow more closely the observed GPP at the site after the observed start of season date than during the time of spring recovery of photosynthesis.

The earliest dates of ENF_SOS_sat occurred in the south-western coastal areas and the latest dates were found in the north-east, ranging from the end of March to the beginning of May (Figure 6). DBF_SOS_sat occurred about one month later than ENF_SOS_sat and ranged from the beginning of May to mid-June. For the regional assessment of JSBACH SOS, we first investigated the spatial agreement between the multi-year averages of ENF_SOS_mod and DBF_SOS_mod for the period 2003–2010 with satellite observations. The modelled mean dates were highly correlated with the respective remote sensing observations (Figure 7, Supplementary Materials: Table S2) showing, however, an early bias for ENF_SOS_mod and a late bias for DBF_SOS_mod. The RMSEs were five and six days for ENF_SOS_mod and DBF_SOS_mod, respectively. For DBF_SOS_mod, this is a bit larger than the accuracy of the remote sensing estimates when considering the site-means (Table 2).

For the boreal sub-regions, the highest spatial correlations between the mean start of season dates for the period 2003–2010 was observed in the middle boreal zone (R² values ≥ 0.79) for both, ENF_SOS_mod and DBF_SOS_mod, while R² was low in the northern boreal zone (<0.5). The bias varied regionally; it increased for ENF_SOS_mod and decreased for DBF_SOS_mod from the southern to the northern boreal zone (Supplementary Materials: Table S2). For ENF, it was largest (≥10 days) in the south-western coastal areas (hemiboreal zone, late bias of ENF_SOS_mod) and south-western Lapland (early bias of ENF_SOS_mod) (Figure 8a). It has to be noted that the satellite estimate based on FSC may be less reliable for the hemiboreal zone in winters with low snow depth [33].

Only a few grid cells in the hemiboreal zone and the most northern grid cells showed a larger bias than 10 days for DBF_SOS_mod (Figure 8b) and the bias was smallest for the transient region between the middle boreal and northern boreal zone.

Throughout Finland, the multi-year average ENF_SOS_mod took place 20–40 days earlier than the simulated date of the JSBACH start of “growth phase” of ENF. The time lag between these two phenological events decreased from south-west to north-east and increased again for northern areas of Lapland.

In the second step, we looked into year-wise comparisons of the start of season dates (Figure 9, Supplementary Materials: Table S3). The correspondence between the annual dates from JSBACH and from satellite observations showed large variations in R², RMSE and bias. As shown in Figure 9a,c, for many adjacent grid cells ENF_SOS_mod occurred on the same day (clumping), whereas we found a more gradual change from south to north in ENF_SOS_sat. In order to assess the reasons for clumping, we analysed the change in 5-day mean temperature, which is a good predictor of spring recovery in ENF, according to Suni et al. [18], and the 5-day mean temperature often exceeded zero degrees concurrent with ENF_SOS_mod. Thus, a change in weather pattern may simultaneously trigger the start of season in a large area. Clumping was not observed for DBF_SOS_mod (Figure 9; lower panel).

The application of the method used for detection of ENF_SOS_sat (Section 2.5.2) relating to the JSBACH-simulated snow cover was not possible for many years (2005, 2007, 2008 and 2009) in the hemiboreal, southern boreal zone and partly also in the middle boreal zone, because the simulated scaled FSC was below 1 during the whole spring period. For the remaining areas with sufficient snow cover, the modelled decrease of FSC occurred in mean earlier than ENF_SOS_sat (8 days) for the period 2003–2010, which is also confirmed by local comparison to snow depth observations in Sodankylä (Figure 4c and Figure 5c).

We verified whether the driving temperature contributed to the differences between DBF_SOS_sat and DBF_SOS_mod. Correlations between the multi-year average dates of the start of the “growth phase” of DBF for the period 2003–2010 from LoGro-P simulations with two temperature drivers (regional climate model and FMI gridded temperature) were very high and the remaining bias for the whole of Finland was very small (R² = 0.98, RMSE = 1.66, Bias = −0.05 days). Nevertheless, the bias varied spatially (Figure 8c); late biases (1–2 days) occurred in central Finland and early biases were observed in north-eastern Finland (−2–−4 days).

The last part of this Section focuses on the inter-annual variations in the start of season dates. On the whole, inter-annual variations of JSBACH model-derived dates were in agreement with satellite observations (Figure 10), except in 2007, when DBF_SOS_mod was 8 days delayed compared to DBF_SOS_sat. The deviation from the respective mean date (period 2003–2010) at individual locations (grid cell) ranged in average between −14 and 10, and −12 and 8 days for ENF_SOS_sat and ENF_SOS_mod throughout Finland. Regarding DBF_SOS, inter-annual variations at one location were slightly smaller than for ENF_SOS, ranging in mean from −8 to 9, and −8 to 7 days for DBF_SOS_sat and DBF_SOS_mod, respectively. The inter-annual variations of mean air temperature for Finland in May correlated with the inter-annual variations in DBF_SOS_mod; that is, higher May temperatures lead to an earlier start of season, (R² = 0.71). Instead, for the evergreen forest, April temperatures are more important: the mean air temperature in April was correlated with the ENF_SOS_mod time series (R² = 0.56). The inter-annual differences between DBF_SOS_mod and ENF_SOS_mod can be largely explained by the differences in the mean air temperature in April and May (R² = 0.70).

We observed an early anomaly in the start of season in ENF in 2007, both from satellite observations and JSBACH simulations (Figure 10 and Figure 11c,f). The yearly anomaly for ENF_SOS_sat was larger than −20 days in south-western coastal areas and in central Finland (Figure 11c). The ENF_SOS_mod anomaly agreed in many parts of Finland with the satellite-observed anomaly, but not for western areas at Bothnian Bay and areas in the northern boreal zone (Figure 11f).

Furthermore, agreement was found for parts of central and northern Finland regarding the early anomaly of DBF_SOS in 2006 (Figure 12b,e); although DBF_SOS_mod indicated a stronger early anomaly than shown by DBF_SOS_sat. Some disagreement between the anomalies was observed in 2007; an early anomaly (up to 15 days) occurred in the south-eastern parts of Finland, according to DBF_SOS_sat, but this was not the case for DBF_SOS_mod (Figure 12c,f).

The accuracy of DBF_SOS_sat was about one week for Finnish sites, which is in line with results obtained in Siberia [70]. Moreover, our observations are in agreement with a map of satellite-derived greening-up of vegetation in Fennoscandia for the period 1982–2002 by Karlsen et al. [77] Figure 2a, but DBF_SOS_sat occurred slightly later in the northern boreal zone than in the map for a relatively warmer period (2000–2006) [28] Figure 5. The sites Aulanko, Kevo, Saariselkä and Värriö were also used for evaluation of the NDVI-based method in studies by Karlsen et al. [28,78] that showed higher correlation between field observations and the NDVI-based start of season in Aulanko and Kevo, but larger RMSE and bias for all four sites than in our results (Table 2).

For the evaluation of DBF_SOS_sat, we compared visual observation of bud break of birch on a few trees with the satellite signal on green-up for many pixels covering an area of several kilometres. As described by Badeck et al. [67], several factors could cause deviations between the in situ observation and DBF_SOS_sat: (a) the inter-individual variability in bud break, (b) systematic time lags between bud break and the date derived from the satellite signal and (c) the heterogeneity in terrain and land cover. Concerning (a), no information on the observer error was given with the ground observations, but the visual observations at Finnish sites were carried out twice a week and we therefore assumed an observers error of about three days. The within-population variation in the date of bud break can range between a few days to up to four weeks and varies from year to year depending on weather patterns [79]. The variation of microclimate for areas covering several kilometres and elevation differences lower than 50 m was estimated with 2.5 days in Badeck et al. [67]. Regarding (b), satellite-observed start of season of DBF did not show a systematic time lag to bud break observations in this study for the whole of Finland. According to Delbart et al. [27,70], the timing of the NDWI minimum in spring after snowmelt corresponds to the time of greening-up of vegetation. Because of noise in the satellite time series, due to cloud cover and atmospheric influences, a threshold value was applied to avoid false detection of the start date (see Equation (3)). We minimized the effect of land cover heterogeneity (c) by selecting MODIS pixels with a dominance of deciduous and mixed forest cover. However, differences in forest structure and the contribution of the understorey to the satellite signal may be one reason for the different biases in DBF_SOS_sat when looking at the southern and northern boreal sites separately (Section 3.1). More investigations are needed on a possible inherent regional bias in the remote sensing estimate for the start of season date in deciduous broad-leaf forest. Information on the forest structure and the understorey phenology, e.g., from web-camera observations, could be helpful in this respect.

In this study, we use the timing of the beginning of snow melt as a proxy indicator for the start of season in ENF that showed good correspondence with ENF_SOS_EC for flux measurement sites in Finland. The quality of the estimate has been discussed in Böttcher et al. [33]. Recent studies suggest that the photochemical reflectance index (PRI) can be used as a direct indicator of spring photosynthetic activation in boreal ENF [25] and can potentially be applied to satellite observations, e g., MODIS [80]. However, satellite-based monitoring with the PRI will need to take into account, among other things, the varying background [25], and this is especially relevant for the spring recovery in evergreen boreal forest when the ground is still snow-covered. Alternatively, the Plant Phenological Index (PPI) seems insensitive to noise and snow and follows well the seasonal development of GPP in boreal ENF [81]. In addition, new space-borne instruments, such as those for the Global Ozone Monitoring Experiment (GOME-2), the Greenhouse Gases Observing Satellite (GOSAT) and the Orbiting Carbon Observatory-2 (OCO-2), provide the opportunity of measuring solar-induced fluorescence and, therefore, ways to directly assess photosynthetic activity [26,82,83], and could be considered in future work.

The presence of long and frequent periods of cloud cover increases the uncertainty of the phenological dates extracted from optical satellite observations. In this study, we applied temporal interpolation of the daily FSC and NDWI time series, but excluded cases of continuous cloud cover longer than two weeks. Including the spatial dimension in the gap-filling based on geostatistical techniques [84,85,86] could improve the completeness of the time series. The quality of the gap-filled time series may still depend on the pattern of missing data in time and space [85]. Another possibility would be the use of the all-weather capabilities of satellite radar and microwave observations alone [87] or in combination with optical observations. For ENF, our method relies on observations of snow cover. Microwave retrievals of the Snow Water Equivalent (SWE), such as provided by the GlobSnow project [88], could be used in combination with the optical observations to gain information on the snow during periods of cloud cover. Currently, the SWE time series covering the period from 1979–2012 is only available at a grid resolution of 25 km × 25 km, but a higher resolution SWE product (augmented using the optical FSC data) is under development at the FMI.

The inspection of the modelled GPP time series (Figure 4 and Figure 5) showed that the JSBACH model instantaneously responds to the increase of air temperature above zero °C resulting in a higher GPP than obtained from the CO₂ flux measurements and, consequently, earlier ENF_SOS_mod than ENF_SOS_EC. The photosynthesis module in JSBACH has the same temperature response for parameters describing the potential electron transport rate and maximum carboxylation rate for the whole year, thus indicating that the simulated vegetation is ready to immediately photosynthesize in spring with increasing air temperatures without any recovery period required, providing there is foliage present as in the case of ENF. However, the boreal coniferous forests have a recovery period in spring after the winter dormancy before they reach their full summertime photosynthetic capacity [89,90]. This may explain the differences between observed GPP at Sodankylä and simulated GPP using local weather observations (Figure 4 and Figure 5). Overestimation of the spring carbon uptake (a too large GPP) by JSBACH was also found by Dalmonech and Zaehle [21] in a study that evaluated, among others, the phenology of the JSBACH model. Furthermore, because of generally lower air temperatures from the regional climate model than from the site measurements, the ENF_SOS_mod,loc was earlier than ENF_SOS_mod.

The observed early bias of three days in ENF_SOS_mod from JSBACH is relatively small in comparison to results by Richardson et al. [14] for the simulated start of photosynthetic uptake from 14 biosphere models (mean bias of −11 ± 15 days) and is in the same range as for the best-performing models in the evaluation. The increase in the negative bias of ENF_SOS_mod from south to north (Figure 8, Supplementary Materials: Table S2) found here, agrees with higher temperatures observed at time of ENF_SOS at northern sites compared to southern boreal sites [4,18,33] and suggests that the response of air temperature for ENF follows a climate gradient in the boreal zone.

For the performance of the JSBACH model regarding the estimation of ENF_SOS, the photosynthesis model plays the dominant role and deviations from ENF_SOS_sat are not due to the LoGro-P model. Improvements to the model performance during the start of season of ENF could be achieved by implementing seasonality to temperature response of JSBACH photosynthesis parameters, and possibly also accounting for photoinhibition in coniferous forests, a phenomenon observed during spring at high light levels and freezing temperatures [91]: the light receiving apparatus in chloroplast gets damaged at high light levels, and for low temperature conditions the photosynthetic capacity may be notably reduced. The description of effects of soil water freeze on carbon uptake could also be improved. According to observations [18,92], the coniferous trees are able to photosynthesize before the soil thaw, but lack of liquid water from soil soon reduces the carbon uptake. In the model, the actual simulated proportions of liquid and frozen water in soil layers could be used to limit the carbon exchange. There are advances related to expressing frozen soil in JSBACH northern permafrost regions [93].

The modelled FSC by JSBACH does not serve as a reliable predictor of ENF_SOS (Section 3.3), as too little snow and too early snow melt is simulated for our study area. Having bias-corrected temperature and precipitation data does not guarantee precise snow cover, as the formulation of snow accumulation relies on a universal temperature threshold and parameterizations of snow processes are not adjusted for region and drivers.

In contrast to results for ENF, we found a late bias of DBF_SOS_mod (Figure 7, Supplementary Materials: Table S2). This agrees with findings by Dalmonech and Zaehle for southern Finland [21], but due to the higher temporal resolution of the remote sensing data set used here, we can quantify the bias better and also give results for the northern parts of Finland that were masked out in the study above. As regards other biosphere models from the study by Richardson et al. [14], the bias for JSBACH DBF_SOS_mod turned out to be relatively small, although direct comparisons are difficult due to differences in the environmental drivers for the model runs.

It has to be noted that the definition of DBF_SOS_mod does not correspond exactly to DBF_SOS_sat. While DBF_SOS_mod is defined based on a minimal change in LAI, the visual observation of bud break may indicate a later stage of development. An optimized parameterization of the LoGro-P model for the boreal zone would allow an improvement of the overall bias for DBF_SOS_mod; variation in parameter values, such as the alternating temperature ( T a l t ) and the minimum value of the critical heat sum ( S c r i t m i n ) can reduce the mean bias (Appendix A, Table A1 and Table A2). A proper tuning of the LoGro-P parameters for our area of interest was outside the scope of this study and needs further investigation, but could be performed using satellite observations and Bayesian optimization techniques.

Similar to ENF, the bias in DBF_SOS_mod showed a regional gradient; that is, the bias decreased from south to north (Figure 7, Supplementary Materials: Table S2). According to field observations, the heat sum requirement for bud break of birch decreases from south to north in the boreal region [94,95] and also the temperature threshold for the NDVI-based greening-up decreases [28,77]. This seems to be in contradiction to our results for JSBACH DBF_SOS_mod. The same critical temperature and heat sum requirement was used in the LoGro-P model for the whole study area. According to the above, one would expect a gradient with earlier bias in southern areas compared to northern areas. We observed the opposite, except for the northernmost parts of the study area (Figure 8b). Reasons for this could be an insufficient chilling predicted by the LoGro-P model, a regional gradient in bias in the bias-corrected driving temperature, or in the remote sensing estimate of DBF_SOS_sat. Under the current climate, the chilling requirement is typically fulfilled in the boreal region [96] and this was also the case in our study area. Therefore, the chilling requirement cannot be used for a better calibration of the south–north gradient, and since the start date of the chilling decay follows a climate gradient, an increase in the decay parameter ( C d e c a y ) would further pronounce the delay in the DBF_SOS_mod in southern regions. The S c r i t m i n parameter plays a role as well, but while the overall bias could be reduced with a lower parameter value, the regional gradient would still remain and this may suggest that optimal parameters for the boreal region are zone-specific.

Despite a very small temperature bias in the start date of the “growth phase” of DBF for the whole of Finland (Figure 8c), the slope of the regression line between DBF_SOS_sat and the modelled date from LoGro-P was slightly higher (0.9 compared to 0.79) when FMI gridded temperatures were used instead of temperature data from the regional climate model. This indicates a small influence of the driving data to the regional gradient. Likewise, the late bias in DBF_SOS_sat for phenological sites in northern, and early bias in southern Finland (Section 3.1) need further investigations and may partly explain differences in bias in JSBACH DBF_SOS_mod in the different boreal sub-zones.