Forest 2200 nm/660 nm reflectance seasonality for improved satellite-image atmospheric correction

The target area, in south and south-east Estonia, lies between longitudes 24.5° E and 28.3° E and between latitudes 57.5° N and 59.6° N. The ground-surface topography is flat, with ground elevation rising only to 318 m above mean sea level (Arold, 2005). The landscape in this part of Estonia is characterised by hills with a local relative height exceeding tens of metres (Arold, 2005).

The forests in the test area are predominantly composed of evergreen coniferous Norway spruce (Picea abies (L.) H. Karst.) and Scots pine (Pinus sylvestris L.), and of broadleaf deciduous European aspen (Populus tremula L.), silver birch (Betula pendula Roth), downy birch (Betula pubescens Ehrh.), grey alder (Alnus incana (L.) Moench), and black alder (Alnus glutinosa (L.) Gaertn.) (Valgepea et al., 2023). Within the test area, diverse forest site types and plant species compositions are developed over various mineral soils (Leptosols, Cambisols, Luvisols, Albeluvisols, and Podzols), and additionally over Histosols (Kõlli et al., 2004). According to Valgepea et al. (2023), the Estonian National Forest Inventory estimates a forest cover of 53.5% for Estonia as a whole.

Forest stand height in the test area generally remains below 35 m (Tappo, 1982). Multi-layer forests are common, with Norway spruce and other shade-tolerant species growing below the upper canopy. The average stand size is 1.24 ha. Within the test area larger stands occur more frequently in state forests (Valgepea et al., 2023). After stand-replacing disturbances, tree regeneration develops dense canopy covers, usually during the first 10 years. In the case of managed forests within the test area, canopy cover may be modified with thinnings and maintenance fellings.

The test area exhibits a diverse range of landscapes, shaped by legal restrictions on clear-cut sizes – ranging from 2 to 7 ha depending on forest type (Forest Act, 2023) – and by a long-standing tradition of retention forest management (Gustafsson et al., 2012) that promotes rapid regeneration on forest land. Within Estonia as a whole, forest management is limited or entirely prohibited over a remarkable 30.3% of forest land (Raudsaar et al., 2023).

The four seasons characteristic of the study area drive its vegetation growth (Jaagus & Ahas, 2000). While observable signs of springtime greening occur in the forest understorey vegetation by mid-April, the active development of new leaves on trees starts by mid-May. The summertime season peak is attained by the end of June, after which growth continues until some point in the first 10 days of September, when the first occurrences of leaf yellowing become evident. By the end of October, the deciduous trees have shed most of their leaves. The timing of the phenophases is spatially variable in Estonia, over a range of about 10 days (Jaagus & Ahas, 2000). The phenophases of forest trees are influenced by tree genotype, and even more by year-to-year differences in air temperature (Possen et al., 2014), which in this region correlate with the North Atlantic oscillation (NAO) and Arctic oscillation (AO) indices (Aasa et al., 2004).

Scots pine in Estonia starts to develop current-year shoots around DOY 132 (EE, 2025). Scots pine has an intensive needle-shedding period in the middle of September, after which only needles from the last two years remain over the winter (Kollist, 1967).

Norway spruce exhibits two well-known phenological forms in Estonia (Sellin et al., 2024), with the average DOYs for budburst on branches cited as 142 and 153 for the early- and late-flushing trees, respectively, and budburst of top shoot few days later. Some phenological observations, however, indicate a somewhat earlier time, with DOY = 138 (EE, 2025). The average DOYs of top-shoot growth cessation are 185 and 190 in the early- and the late-flushing spruces, respectively, whereas the respective DOYs for growth cessation of branches are 175 and 185. In contrast with Scots pine, Norway spruce needles undergo no massive, visually detectable, autumnal yellowing.

European aspen likewise exhibits early- and late-flushing phenological forms in Estonia. Foliation starts around DOY 132, can in some years be delayed by a week or two, and lasts for about 14 days (Tamm, 2000). Autumn colouring appears in the first leaves in September, becoming fully evident by DOY 266 (EE, 2025). Dormancy is reached on DOY 289 (EE, 2025).

Two birch species, silver birch and downy birch, are included in our forest inventory data without being distinguished. Downy birch dominates on poor soils, whereas silver birch occupies most of the fertile soils. On average, birch leaf formation starts on DOY 126 (Jaagus & Ahas, 2000), or according to the phenology calendar (EE, 2025) a few days later, on DOY 130. Our own observations indicate that leaf unfolding is delayed by about a week on wet and poor soils, since these warm more slowly, hindering the start of active growth. Autumn colouring starts in the first leaves by the end of August, becoming fully evident by DOY 262. Dormancy is reached on DOY 295 (EE, 2025).

Grey alder is described in the phenology calendar (EE, 2025) as follows: bud swelling starts on DOY 109, autumn leaf colouring is fully observable on DOY 271, and dormancy is reached on DOY 262. Our own observations indicate that autumn colouring of grey alder foliage yields brownish tones, in contrast with the yellow of silver birch and the yellow-red mixture of European aspen.

The Estonian Environment Agency (Keskkonnaagentuur) maintains the Estonian forest register (Eesti metsaregister). The database contains forest management inventory data in tabular form, accompanied by 1:10,000 stand border maps. We used state forest records taken from the database in the version of 08 February 2023 to select stands that are sufficiently large in relation to 30 m pixels in satellite images. Our database query conditions were as follows: stand age greater than 15 years; stand area between 1 and 10 ha; and relative stocking density (this correlates strongly with canopy cover) greater than 80%, including both the upper and second tree layer.

We made the selection from the list of stands dominated by Norway spruce, Scots pine, European aspen, birch species, and grey alder. Black alder stands were not included, because preliminary experiments with the satellite-image atmospheric correction algorithm revealed black alder stands to be not selected as dense dark vegetation.

For the analysis of seasonal changes in ρ_2.2/ρ_0.6, the selected forest stands were assigned tree species group labels, the various groups being defined by soil fertility and dominant tree species (Table 1). The final selection of stands was made only after queries from raster layers in the satellite images, to ensure that only those stands were included that had more than 70% of their area covered by pixels with centroids inside the stand polygons. This exclusion rule eliminated stands of irregular shape and minimised the influence of mixed pixels located at stand borders. Table 2 gives sample sizes and mean values of forest inventory variables. Norway spruce contributes notably to species composition in the deciduous broadleaved group (9%) and in the pine stands growing on fertile soils (7%), because Norway spruce is shade-tolerant, frequently growing as a second layer below the light-demanding deciduous broadleaved species, and also frequently growing on fertile soils below stands of Scots pine. The mean normalised difference vegetation index (NDVI) in our selected forest areas is 0.79, with standard deviation of 0.09 (Table A3.1), when calculated for the active growth season (140 < day < 260).

Landsat-8/9 OLI (Zanter, 2016; Sayler & Glynn, 2022) collection 002 images that were processed to level L1 (at-sensor radiance) and L2 (surface reflectance) were downloaded from the USGS GLOVIS data server. The Sentinel-2A/B MSI images of the latest available baseline processing with versions starting from 04.00 (https://sentiwiki.copernicus.eu/web/s2-processing) were downloaded from the Estonian Land Board (Maaja Ruumiamet) data hub. The Sentinel-2 MSI data corresponded to at-sensor reflectance at the processing level L1C.

An inspection revealed that the USGS Landsat-8/9 OLI surface reflectance dataset was influenced by a production error that expressed itself as a noise pattern deriving from the aerosol estimation quality map in the red spectral region and even in the nearinfrared. For this reason, USGS Landsat-8/9 OLI surface reflectance data (L2) were not used for forest stand-level analysis, but only for determining the ρ_2.2/ρ_0.6 mean value across each image. The image-wide dark dense-vegetation mean reflectance values were used only to estimate the initial parameters for our ρ_2.2/ρ_0.6 phenology model: USGS applies a phenology-dependent algorithm (Vermote et al., 2016) for atmospheric correction, in contrast with the Sen2Cor3 algorithm (Pignatale et al., 2025) employed in this study to process both Landsat-8/9 OLI L1 and Sentinel-2 MSI L1C data.

In the case of Landsat-8/9 OLI, this study used images from WRS-2 system paths and rows 185–019, 185–020, 186–019, 186–020, 187–018, 187–019, and 187–020. In the case of Sentinel-2 MSI, this study used image granules from the locations of MGRS-2 system tiles T35VME and T35VNE. Our decision in both cases was driven by the availability of stationary AERONET (https://aeronet.gsfc.nasa.gov) and SkySpec (Kuusk & Kuusk, 2018) radiometers that provided data for the preliminary assessment of atmospheric correction of satellite images, in work parallelling Doxani et al. (2023). However, when our study progressed beyond this preliminary assessment, we did not find it necessary to analyse the statistical correlations between the radiometer data and aerosol optical depth estimates from the satellite images.

In selecting a satellite image for the study, we required that the usable area, net of obscuration by clouds, be greater than about 25% of the image area. The sampling period started each year in March, with the consequent inclusion of images that may on a local scale have exhibited some remaining small traces of snow in shaded pixels at forest edges. Each year, the sampling period ended on the last day of October. In the process of image selection, the entire time series, starting in 2013 and ending in 2024, was checked visually. The number of days when the overpass of at least one satellite yielded usable image data was 268. This total, however, includes days when both satellites made overpasses of the study area: in the entire sequence of selected images, there were 122 Landsat-8/9 overpasses and 146 Sentinel-2 overpasses.

All satellite images were used in their original map projection. Forest stand boundaries were reprojected for each individual image to avoid errors from the resampling of raster data. High-quality cloud and cloud-shadow masks were constructed for the Sentinel-2 MSI images with KappaMask (Domnich et al., 2021). Our assessment supplied additional confirmation that this algorithm has substantially greater precision than other algorithms. In the case of Landsat-8/9 OLI, the cloud masks bundled with the spectral data were later additionally used in the Sen2Cor3 runs, as the masks were superior to the native Sen2Cor3 algorithm in detecting faint clouds and faint cloud shadows.

The Sen2Cor3 (Pignatale et al., 2025) tool is a version of the official atmospheric correction program Sen2Cor distributed by the European Space Agency and maintained by its partners. Sen2Cor is deployed worldwide in data hubs and by Sentinel-2 MSI data users for predicting ground surface reflectance from top-of-atmosphere (TOA) reflectance measurements (ESA, 2021). Sen2Cor3 is developed within the Sen2Like project (Saunier et al., 2022) and includes an option to process also Landsat-8/9 TOA data and calculate surface reflectance.

We downloaded the latest Sen2Cor3 version 3.01.00, available in August 2024, according to the instructions given in GitHub (https://github.com/senbox-org/sen2like/blob/master/sen2cor3/README.md), installing it on a Linux workstation. This software, written in Python, is provided as open-source code, apart from the module for calculating topographic shadows.

Initial test runs indicated the possibility of improving the preprocessing phase, run by Sen2Cor3 as a preliminary to its atmospheric correction. In this preprocessing, Sen2Cor3, proceeding on the basis both of reflectance ρ_2.2,ref in the near-infrared (2.2 μm) band and of other filters, selects pixels that are assumed to represent dense dark vegetation (DDV). First, we replaced the three-step adjustment of threshold imposed on a preliminary estimate of ρ_2.2,ref by Sen2Cor3 in its DDV-pixel selection phase with a small-step incremental search for the threshold value. Our search used the same three configurable decision thresholds as are provided with the program configuration files described in the Sen2Cor3 manual. In place of the commonly used value ρ_2.2,ref = 0.05, however, we set the first threshold to the lower value of 0.035.

Second, we added a binary morphologybased filter to remove small putative DDV-pixel groups that were found upon visual inspection of the DDV maps to consist mostly of noise. We constructed our filter by combining SciPy library procedures that do binary erosion, setting the parametric window structure size to 2 × 2 pixels, and a binary propagation operation. We found that our filter effectively removed salt-and-pepper noise from the DDV map without generating the geometric distortions typical in the case of moving window-based median filters.

This pair of enhancements yielded a large improvement in the consistent detection of DDV pixels, as the target optimum of 2% out of the total number of pixels of interest: the DDV set was now more efficiently targeted, without the occasional large incremental steps in the DDV-pixel set size generated by the original algorithm.

However, test runs under our two enhancements now revealed that DDV pixels were frequently selected from areas of cloud shadow not identified by the Sen2Cor algorithm. This problem was in turn addressed with a third modification, by importing data for clouds and cloud shadows from KappaMask, as applied on the one hand to the Sentinel images, on the other hand to the Landsat L1 pixel-quality metadata layer bundled with images in the Landsat product. Our selection of DDV pixels was allowed only for terrain lying outside the buffers of 600 m for Sentinel-2 MSI and 1500 m for Landsat-8/9OLI.

Our three enhancements jointly added about 10 seconds of processing time per image tile to the total of 15 to 25 minutes originally required.

As a filter throughout its processing, Sen3Cor uses surface-elevation maps (for the calculation of vertical visibility in the atmosphere) and water-body maps. In place of the coarse-scale datasets included within the Sen2Cor3 installation package, we used the ground-surface digital elevation model, at fine-scale (25 m) spatial resolution, with a 1:10,000 water body map (showing lakes and rivers) provided by the Estonian Land Board. To avoid the selection of mixed-landcover pixels as dark dense vegetation, the water body map was rasterised, and a buffer was added around water bodies.

The atmospheric correction procedure requires data for ozone and water vapour content. In the case of Sentinel-2 MSI, the ozone data is bundled with the image, and a water-vapour map is predicted with an internal algorithm of Sen2cor3. In the case of Landsat OLI, the Copernicus Atmosphere Monitoring System (CAMS) data (https://atmosphere.copernicus.eu/) could be used. Although this functionality was not yet fully included in the Sen2Cor3 version 3.01.00, the Telespazio team kindly provided us with code from the development version of the next Sen2Cor3 release. This Sen2Cor3 upgrade allowed us to read ozone and water vapour data from daily CAMS data, instead of setting some values manually for each individual image.

Finally, we added functionality for setting the target region of interest (ROI) for DDV-pixel selection, through a vector polygon file. We constructed this ROI by adding a 30 km positive buffer around the Estonian state border, thereby excluding from the images localities contaminated by haze outside our study area. Because hazy atmospheres exhibit trends in their optical properties, the imaging of such contaminated localities would bias otherwise usable pixels. Specifically, bias would be introduced by the fact that Sen2Cor uses mean aerosol optical depth per granule instead of the map (for reasons yet unknown to us: the algorithm duly constructs the map in its internal operations, while leaving it unused).

The processing of the entire image time series took on the order of 12 or 13 hours on a Linux workstation running the OpenSUSE 15.5 operating system, with 16 CPU cores, 128 GB of RAM, and a NVIDIA GeForce 630 graphics card fitted with its own 2 GB of RAM. System tools were used to run from 12 to 14 Sen2Cor3 instances in parallel, with each instance processing a single image granule.

With the enhanced Sen2Cor3 all images were first processed under standard settings, except that the initial threshold for ρ_2.2,ref was specified as 0.035 in place of the standard value 0.05. The ρ_2.2/ρ_0.6 ratio was kept at the constant value 2, corresponding to the value 0.5 given in the program input-configuration file. This first run produced DDV-pixel maps.

The DDV maps were then used to sample pixels from the USGS version of the Landsat-8/9 OLI surface-reflectance data. This sampling step was appropriate because the LaSRC algorithm (Vermote et al., 2016) uses coarse-resolution reflectance data time series collected from a selection of MODIS and MISR images (with pixel size of about 5 km × 2.5 km at our study area) to determine band ratios for particular regions and seasons: consequently, the USGS Landsat-8/9 OLI level L2 dataset, for which LaSRC considers the seasonality of ground-surface spectral reflectance, should in theory outperform Sen2Cor in capturing the seasonal changes in plant canopy reflectance. The mean ρ_2.2/ρ_0.6 ratio was calculated for each Landsat-8/9 OLI granule, and to this data the following modified double logistic model (Eklundh & Jönsson, 2012) was fitted: 1ρ2.2/ρ0.6=a+1b+exp(kpk−tkk)−cd+exp(kps−tks)+ε,{\rho _{2.2}}/{\rho _{0.6}} = a + {1 \over {b + \exp \left( {{{{k_{pk}} - t} \over {{k_k}}}} \right)}} - {c \over {d + \exp \left( {{{{k_{ps}} - t} \over {{k_s}}}} \right)}} + \varepsilon , where ε is the model residual error (RSE); t is the day number in the year; a, b, c, and d are free parameters; k_pk is the springtime inflection point; k_k is the springtime growth rate; k_ps is the autumn inflection point; and k_s is the autumn growth rate. The values of k_pk, k_k, k_ps, and k_s were first found by expert guess, after which a, b, c, and d were estimated with the nls procedure in R (R Core Team, 2022; Baty et al., 2015), version 4.2.2. The obtained model estimates, with 100 degrees of freedom and the residual standard error 0.087, were as follows: a = 1.81, b = 6.51, c = 0.379, d = 0.785, k_pk = 140, k_k = 5, k_ps = 260, and k_s = 25.

Next, the model was fitted to the surfacereflectance data that had been obtained from the first run of Sen2Cor3 (described above), with its assumption of constant ρ_2.2/ρ_0.6 = 2. The empirical dataset for this fitting step consisted of DDV pixels aggregated (binned) for each image by unique ρ_2.2 values, to decrease the data size. The group mean ρ_0.6 was calculated for each bin, with the number of pixels in the bin used as a weight in model parameter estimation. The parameters k_pk, k_k, k_ps, and k_s were fixed to the values found for the USGS Landsat L2 dataset. The free parameters a, b, c, and d were again estimated with the nls procedure in R. After this revision, the model estimates, with 168,484 degrees of freedom and the residual standard error 9.4, were as follows: a = 2.08, b = 6.85, c = 0.40, d = 1.1, k_pk = 140, kk = 5, k_ps, = 260, and k_s = 25. The notable large residual error arose because here the DDV pixel groups were the observations.

With the model (1) predictions checked against data scatter plots, the model was found suitable. The model was then used with the adopted parameter estimates to make predictions of ρ_2.2/ρ_0.6 for all the images in our dataset. Atmospheric correction of top-of-atmosphere radiance data was done again with Sen2Cor3, but now dropping the assumption of the constant channel ratio and using instead the DOY-dependent ρ_2.2/ρ_0.6.

Finally, the atmospherically corrected images served to sample forest spectral reflectance, using forest stands as delineated in the forest inventory map. The empirical dataset for studying ρ_2.2/ρ_0.6 seasonal variability was constructed by selecting the stands that had (as previously described) more than 70% of their area covered by reflectance data pixels with centroids inside the stand polygons. An additional filter was now imposed, excluding those areas in each image that were less than 1 km away from clouds and cloud shadows.

At this stage the model (1) was fitted for each tree species group (Table 1, Table 2), so that single observation corresponded to the ρ_2.2/ρ_0.6 of an individual forest stand calculated from a particular image granule. The search ranges for the model parameters k_pk, k_k, k_ps, and k_s were allowed to vary reasonably widely around the previous estimates. The search ranges for a, b, and d had to be limited for some species groups. The parameter c was throughout this fitting fixed to the constant c = 1. The coefficient of determination R² was calculated as 2R2=∑ (yi−y^i)2∑ (yi−y¯)2,{R^2} = {{\sum {{{\left( {{y_i} - {{\hat y}_i}} \right)}^2}} } \over {\sum {{{\left( {{y_i} - \bar y} \right)}^2}} }}, where y_i is the ρ_2.2/ρ_0.6 value for a single stand on an image granule, ŷ_i is the predicted value, and ȳ is the species-group sample mean. The model (1) was incapable of following some changes characteristic of autumn. Empirical values were consequently also approximated by splines, using the smooth splines procedure in R, and adjusting the degrees of freedom to make the fitted spline follow the main trend through the composite year without picking up too much variability within short time ranges.

In all our twelve species groups, the starting value of ρ_2.2/ρ_0.6 after the melting of snow in March and during the first week of April is found to remain between 2.2 ≤ a ≤ 2.4 (Table 3). This is greater than the commonly assumed constant ρ_2.2/ρ_0.6 = 2. During the first ten days of April, the landscape-level ρ_2.2/ρ_0.6 value starts to increase (Figure 1), rapidly reaching an inflection point k_pk around day 140 (Table 3). The maximum values of ρ_2.2/ρ_0.6 occur in the vicinity of the summer solstice, around day 180. This is followed until the end of July by a slightly decreasing near-plateau. In August, an accelerated decrease of ρ_2.2/ρ_0.6 starts. This process continues until the middle of September, by which the ρ_2.2/ρ_0.6 band ratio has decreased almost to 1.8. After this minimum is reached, an increase resumes, continuing until the end of October (the last month in our dataset). The season-ending value is about the same as the starting value.

Figure 1.

The average seasonal course of the red ρ_0.6 and near-infrared ρ_2.2 reflectance ratio, summarised as a boxplot using 5-day bins. The dashed line corresponds to the preliminary model estimated from USGS Landsat-8/9 OLI L2 surface reflectance data. The wide line corresponds to model (1); the narrower line to our spline-based refinement.

However, the seasonal ρ_2.2/ρ_0.6 value for evergreen needleleaf forests is found not to depend so strongly on the time during the season. Although the 5-day averaged values indicate small fluctuations, variability around the mean values in each 5-day bin hides most of the possible seasonal changes, and the double logistic phenology model (1) predicts only a marginal increase for the mid-summer (Figure 2). This independence of the evergreen needleleaf ρ_2.2/ρ_0.6 value from the day of the year is indicated also by the almost zero value of the determination coefficient (Table 3). In contrast to the evergreen needleleaf tree species group, the deciduous broadleaved species group is found to exhibit substantial dependence on the date during the growth season (Figure 3). While there is not much difference in the springtime inflection point k_pk, the autumn inflection point k_ps, of the phenology model occurs 7 to 10 days earlier, in comparison both to the landscape average and to the evergreen needleleaf stands (Table 3). The autumn ρ_2.2/ρ_0.6 minimum is much more pronounced than in the landscape-level average (Figure 3). However, the model (1) is not flexible enough to capture such small fluctuations. On the other hand, the splinebased smoothed model (Table A2.1) follows the phenomena rather well. We find that during the autumn minimum the median values of ρ_2.2/ρ_0.6 can decrease even to about 1.5 for about 10 days.

Figure 2.

Evergreen needleleaf tree species group. Figure details are as in Figure 1.

Figure 3.

Deciduous broadleaved tree species group. Figure details are as in Figure 1.

The fitted parameters for particular tree species and soil fertility groups in model (1) follow the characteristics of the general groups that were defined in terms of pooled leaf type and pooled tree-shoot morphology. The run of seasonal ρ_2.2/ρ_0.6 values for Scots pine stands and Norway spruce stands is found to resemble the seasonal ρ_2.2/ρ_0.6 run for pooled evergreen forest (Figures A1.1 to A1.4). Similarly, the European aspen stands, birch stands, and grey alder stands are found to follow the seasonal ρ_2.2/ρ_0.6 run of the pooled deciduous broadleaved species group (Figures A1.5 to A1.9).

The tree species subsets do, however, exhibit some possible small, interesting time periods during which the empirical values of ρ_2.2/ρ_0.6 do not follow the phenology model (1).

(A) In the autumn, the three Scots pine stand groups and the Norway spruce stand group are found to exhibit a small decrease in ρ_2.2/ρ_0.6, similar to the ρ_2.2/ρ_0.6 decrease in the various deciduous broadleaved species groups. In the first Scots pine stand group (Scots pine overall, without differentiation by soil fertility) the decrease starts possibly already from day 230, with ρ_2.2/ρ_0.6 reaching its minimum value around day 270. In the Scots pine stands growing on poor soils this feature is more pronounced (Figure A1.3). In the Norway spruce stands group, the autumn minimum at the end of September is still more strongly pronounced (Figure A1.4). Each of the latter two groups includes a small proportion of deciduous broadleaved trees (Table 2), the share being about 10% in the case of the Norway spruce stands. This heterogeneity in stand composition partly explains the similarity with the broadleaved deciduous groups. Additionally, however, in the Estonian autumn about 25% of Scots pine needles turn yellow and are shed. This can cause seasonal reflectance changes similar to those exhibited by many broadleaved deciduous species.

(B) In the birch stand groups the seasonal changes in ρ_2.2/ρ_0.6 are found to follow those of the pooled deciduous broadleaved group. The autumn minimum is, however, more pronounced in the stands growing on fertile soils (as can be seen by comparing Figure A1.6 against Figures A1.5 and A1.7). The ρ_2.2/ρ_0.6 value increases remarkably during the growth season, attaining values around 3 for several months. In the European aspen stands the ρ_2.2/ρ_0.6 value increases even more, with 5-day bin medians yielding ρ_2.2/ρ_0.6 ≥ 3.25 (Figure A1.8, Table A2.1). Otherwise, however, European aspen stands and birch stands exhibit similar seasonal values. The grey alder stands reach the greatest values of ρ_2.2/ρ_0.6 among all groups, with 5-day bin medians rising even as high as 3.5, and they differ somewhat in their seasonal courses from birch and European aspen stands (Figure A1.9, Table A2.1). There is in their case no rapid increase phase in the spring, with the ρ_2.2/ρ_0.6 seasonal course instead almost symmetrical around the summer solstice. Nevertheless, the springtime plateau before budburst is detectable, as in the other deciduous broadleaved species groups. An interesting pattern can be found during the late summer period, from day 250 until day 280, when the decrease in ρ_2.2/ρ_0.6 is replaced by a temporary increase before the final decrease sets in.