A total of 15 living P. engelmannii trees (1 core per tree) were selected for dendroanatomical measurements from a collection sampled in 2015, from tree-line sites (2000–2100 m a.s.l.) adjacent to the Athabasca Glacier in the Columbia Icefield area of the Canadian Rockies (52.13∘ N, 117.14∘ W, Fig. 1). The selection of cores was based on (1) the visual appearance of the material (cores with obvious defects were avoided), (2) the temporal coverage of the series (we strived to have an even replication through time) and, (3) the common signal strength of the ring width measurements (in general, cores with higher than average correlation with the master chronology built from all the other series were selected for wood anatomy). The selection was primarily dictated by factors (1) and (2), and only secondarily by factor (3).

Wood cores were refluxed in alcohol for 24 h using a Soxhlet apparatus to remove resin and other soluble substances, and subsequently embedded in paraffin using a Tissue Processor TP1020 and Histocore Arcadia Embedding Center (Leica, Germany). A rotary microtome RM2245 (Leica Biosystems, Germany), equipped with N35 disposable microtome blades (Feather, Japan), were used to cut 12 µm thick transverse sections from the wood cores. The thin sections were stained with a 1:1 safranin-astrablue solution and mounted on slides with Euparal (Carl Roth, Germany), following standard procedures (von Arx et al., 2016). Digital images from each section were taken with a Zeiss Axio Scan Z1 (Carl Zeiss, Germany) at a resolution of 2.3 pixels µm-1. Tree ring borders and individual tracheid cells were then semi-automatically identified, and ring width as well as the position and anatomical dimension of each tracheid cell were measured in the digital images using the image analysis software ROXAS (v3.1) (von Arx and Carrer, 2014). The anatomical parameters included, for instance, cell lumen area and cell wall thickness (CWT), where the latter was measured in four directions to obtain the average cell wall thickness, i.e., two radial and two tangential cell walls per tracheid cell (Prendin et al., 2017). Each tree ring was divided into 20 µm wide bands parallel to the ring border (the tangential extension of each band encompasses ∼ 75–100 tracheids). In order to minimize the influence of outliers, for each anatomical parameter the values of all cells corresponding to the 75th percentile within each 20 µm-wide band were retained for further analysis, i.e., from the cells within the radial and tangential extension of each band, the 75th percentile of each cell dimension was calculated, building up intra-annual measurement profiles of CWT, lumen area, and other anatomical traits. The maximum parameters (e.g., Max. radial CWT) thus retain the highest value of the profile for each year, the minimum parameters retain the lowest value of the profile for each year, and the earlywood and latewood parameters retain the average value of the profile for the earlywood and latewood portions of the ring, respectively. The anatomical density was derived as the ratio of wall area to overall cell area (i.e., including both wall and lumen area) in each 20 µm wide band. Mork's index (Denne, 1989) was used to separate the earlywood and the latewood portions of the ring. For further details regarding the dendroanatomical measurements, see Björklund et al. (2020).

From the potentially large number of possible dendroanatomical parameters, we narrowed down subsequent analyses to seven parameters of anatomical dimensions, and three wood density parameters based on anatomical dimensions, which are directly comparable to X-ray and MXBI-based microdensitometric parameters. The parameters are listed in Table 1. For comparative purposes, we also retained X-ray derived measurements of MXD (Luckman and Wilson, 2005), and the previously unpublished latewood blue intensity counterpart (MXBI) measured on P. engelmannii from the Columbia Icefield area. The X-ray MXD was produced using radiodensitometric techniques (Schweingruber et al., 1978) from 1.2 mm-thick laths, cut using a twin-blade saw along the tree cores but perpendicular to the fiber direction (see Luckman and Wilson, 2005 for details). For the production of MXBI, the methodology outlined in Rydval et al. (2014) was adopted. The MXBI measurements were conducted using the CooRecorder software (ver. 8.1) (http://www.cybis.se/forfun/dendro/index.htm, last access: 12 January 2021). Corresponding time series of ring width were also obtained and hereafter referred to as original ring width, as opposed to ROXAS ring-width, which were measured in the ROXAS program on the 15 cores used for the dendroanatomical measurements. The X-ray MXD and MXBI datasets were originally developed from living trees and snag material; however, to ensure consistency for the parameter comparison, we used X-ray MXD, MXBI and original ring width measurements from living trees only (X-ray MXD: N=78 series, MXBI: N=182, and original ring width: N=182, see Table 1). The dendroanatomical analysis was performed on tree cores for which original ring width and MXBI measurements were available. Thus, an additional subset based on the 15 trees was retained for the latter 2 parameters to also ensure a direct comparison with the dendroanatomical chronologies. For the full MXBI dataset (N=182), we additionally derived eight partly overlapping percentile chronologies based on absolute ring width (Fig. S1 in the Supplement), to assess whether a similar ring width dependence as previously reported by Björklund et al. (2019) from northern Fennoscandia could also be detected in the Icefield dataset, i.e., ring width-related differences of MXBI measurements taken in narrow versus wide rings. The following ring width percentile intervals were used: 0–30th, 10th–40th, 20th–50th, 30th–60th, 40th–70th, 50th–80th, 60th–90th, and 70th–100th to derive the subsampled MXBI chronologies. Thus, for example, the 70th–100th percentile chronology is computed from MXBI values measured in the 30 % widest rings, while the 0–30th percentile chronology corresponds to MXBI values from the 30 % of the narrowest rings. Unfortunately, it was not possible to conduct a similar comparative analysis for the X-ray MXD as the corresponding ring width measurements originally developed were unavailable to us in the current study.

Since the analysis was performed on data derived from a cohort of same-aged living trees, capturing low-frequency variability (i.e., decadal and longer) with regional curve standardization (RCS) type methods is a challenge (e.g., Briffa et al., 1992). This is because living trees only, share potential climate signal on lower frequencies even if they are aligned by cambial age (trend in signal) (Briffa and Melvin, 2011). Any attempt, even using signal-free approaches, will provide indices that likely would have to be revised when implementing the same technique on a large multigenerational dataset, and ultimately reflect certainty where there is little. Thus, we primarily focused here on the year to year signals in the tree ring anatomical parameters. To emphasize the interannual variations, the individual dendroanatomical series were detrended in the program MATLAB (version R2021a), by (1) fitting a cubic smoothing spline function with 50 % frequency response cut-off at 35 years to the raw tree ring series (Cook and Peters, 1981), (2) subtracting the fitted values from the observed values to obtain detrended series (division was used to standardize the ring width measurements), and finally (3) averaging the detrended series by simple arithmetic mean to produce the final parameter-specific chronologies (hereafter referred to as detrended data). The same detrending procedure was performed on the MXBI, X-ray derived MXD and original ring width series in order to obtain data that are comparable with the dendroanatomical datasets. All chronologies were truncated to the 1700–1994 period in the subsequent analyses to ensure a consistent overlap between datasets as well as a sample depth ranging between 9 and 15 cores for the anatomical dataset.

To evaluate the strength of the between-series common signal and establish the replication needed to obtain mean chronologies meeting the commonly accepted standard, we used the RBAR (defined as the mean Pearson's correlation coefficient between all possible pairs of individual tree ring series) (Wigley et al., 1984) and expressed population signal (EPS) (Briffa et al., 1992) statistics. To assess the degree to which the various parameters co-vary, principal component analysis (PCA) and pairwise correlations were computed over the 1700–1994 period.

Detrended tree ring parameter chronologies were assessed for their relationship to regional monthly mean (Tmean) and maximum (Tmax) temperatures, by correlation against the monthly 0.5∘×0.5∘ gridded CRU TS v4.03 dataset (Harris et al., 2020) for the grid point average bounded by the latitude and longitude coordinates 48.25–55.75∘ N/113.75–123.25∘ W (Figs. 1, 2). The Tmax was included in the analysis because previous work has demonstrated slightly stronger calibration statistics than for Tmean when using MXD and ring width chronologies for climate reconstruction in this region (e.g., Wilson and Luckman, 2003; Heeter et al., 2021; Wilson et al., 2019, 2014). The associations with monthly precipitation totals and minimum temperatures were also tested, but not included here due to weak significant empirical relationships. The lack of precipitation sensitivity of P. engelmanni in the Icefield area was already noted in St. George and Luckman, 2001, which is not surprising as the trees are growing in temperature-limited upper tree-line environments. Pearson's correlations were calculated between parameter-specific chronologies and monthly meteorological variables over the 1901–1994 period, and the 1901–1948 and 1949–1994 subperiods to evaluate temporal stability of the climate responses. A paired t test was used to test whether the calibration statistics differed between tree ring parameters and sub-periods. To make the climate sensitivity analysis comparable to previous studies from the Columbia Icefield area, we also included the homogenized (1895–present) 50×50 km gridded temperature data (Zhang et al., 2000; Vincent and Gullett, 1999) originally developed by the Meteorological Service of Canada and previously used in Luckman and Wilson (2005) to reconstruct last-millennium summer temperatures for the Canadian Rockies. Similar to Luckman and Wilson (2005), we used the mean of four grids closest to the Columbia Icefield area. Calibration trials with these data are provided in the supplement (Figs. S2 and S3). To ensure the climate analysis was not affected by long-term trends, all temperature data were filtered prior to analysis using the same 35-year filter as was used to detrend the tree ring parameters (henceforth referred to as detrended data).

Furthermore, the dynamic nature of the temperature signal (i.e., optimal target season and its temporal stability) was evaluated through moving window correlation analysis between detrended tree ring chronologies and detrended average daily temperature data (grid 52.5∘ N, 118.5∘ W) from the Berkeley Earth dataset (http://berkeleyearth.org/data/, last access: 31 January 2021) (Rohde and Hausfather, 2020) covering the 1880–recent period. Pearson's correlations were computed for 30-year sliding windows with a 1-year offset. For each 30-year block, temperatures were averaged in 30 d long windows which were shifted at daily time steps throughout the year (sensu Jevšenak and Levanič, 2018).

Besides the conventional width parameters (i.e., ring width, earlywood and latewood width, referred to as ROXAS in Table 1), seven anatomical parameters and three anatomically-based density parameters were retained for analysis. Basic chronology assessments of detrended data over the common 1700–1994 CE period are provided in Table 1, and non-detrended mean chronologies for selected parameters are shown in Fig. 2. In line with previous work by Björklund et al. (2020) on temperature-sensitive conifers, we found that maximum radial cell wall thickness (Max. radial CWT) and anatomical MXD (aMXD) are the two anatomical parameters with the highest mean inter-series correlation (RBAR = 0.47 and 0.48, respectively). For both parameters, EPS reaches the 0.85 threshold (Wigley et al., 1984) with 6 series (Table 1). Notably, these values are of comparable strength to the RBAR and EPS of X-ray based MXD (RBAR = 0.49, 6 trees required for EPS = 0.85). By comparison, the RBAR for MXBI is surprisingly low at 0.19 and the replication needed to attain the EPS of 0.85 is 24 series. These MXBI chronology statistics are lower than for ring width (RBAR = 0.22 and 0.28 for original and ROXAS ring width, respectively), an observation noted previously by Rydval et al. (2014) and Wilson et al. (2019). The RBAR and EPS values for MXBI slightly decrease if computed only on the 15 trees that have been preselected for the dendroanatomical analysis. This is surprising given that the selection of the cores for dendroanatomy was partly based on its ring width signal strength (see Sect. 2.1), and that the RBAR and EPS statistics for ring width actually improve when narrowing the analyses down to these 15 trees (see Table 1). Although the BI-based density parameters typically require a larger sample size than ring width (e.g., Blake et al., 2020; Wilson et al., 2021) for a robust chronology, the MXBI chronology statistics obtained for P. engelmannii from our site are still lower than the previously reported MXBI findings for the same species across British Columbia, Canada (Wilson et al., 2014).

Notably, several anatomical and density parameters are found to exhibit a relatively low common signal, yet a reasonably strong temperature sensitivity (see Sect. 3.2). These include, in decreasing order of signal strength: earlywood (EW) cell wall area (RBAR = 0.13), EW lumen area (RBAR = 0.12), EW density (RBAR = 0.10), EW cell area (RBAR = 0.09) and latewood (LW) cell area (RBAR = 0.09). The replication required to attain robust EPS statistics ranges between 38 (EW cell wall area) and 57 trees (EW cell area and LW cell area).

The co-variability between the various parameters over the common 1700–1994 period was assessed through principal component analysis and pairwise correlations (Fig. 3). The first two components together represent 68.1 % of the total variation. The PC1 alone explains 43.8 % of variance, and is dominated by latewood-related parameters, including both anatomy and density parameters. We found that aMXD, Max. radial CWT and X-ray MXD cluster together in the bivariate plot, showing that all three parameters express comparable signals (also corroborated by the correlation matrix in Fig. 3b). The MXBI also loads strongly positively on PC1, but slightly separates from this cluster by being positively correlated to PC2. Among the LW density-related components, MXBI is the parameter best correlated with ring width and latewood width chronologies (Fig. 3b), although these correlations are only moderate (rMXBIvs.originalringwidth=0.43, rMXBIvs.latewoodwidth=0.66). The principal component analysis including the subsampled MXBI percentile chronologies based on the absolute corresponding ring widths reveal that the correlation coefficients against the latewood width, and to some degree also ring width, successively increase for the “narrow-ring MXBI chronologies” (Fig. S5). The “wide-ring MXBI chronologies” (i.e., ∼ 50th–100th percentiles) are, on the other hand, more similar to the aLWD, Max. radial CWT, aMXD and X-ray MXD chronologies. This observed ring width inclination of MXBI suggests that the dataset might be subject to a resolution bias (Björklund et al., 2019). More detail on this issue is given in Sect. 3.4.

The variance of PC2 (24.3 % of total variability) is dominated by ring width and earlywood-related density and anatomy parameters. Amongst these, EW density stands out by loading strongly negatively on the PC2 axis (reflecting its negative association with early summer temperatures, see Sect. 3.2). Moreover, the EW cell wall area stands out by loading more strongly on the PC1 axis than on the PC2 axis, and by clustering more closely with the latewood than with the earlywood components (reflecting its late summer temperature sensitivity, see Sect. 3.2).

Simple linear correlations between selected parameters and monthly CRU TS mean (Tmean) and maximum (Tmax) detrended temperatures are shown in Fig. 4. In line with previous work from North America (Heeter et al., 2021; Wilson and Luckman, 2003; Luckman and Wilson, 2005; Wilson et al., 2014; Harley et al., 2021), our results reinforce the importance of Tmax temperatures for wood formation and growth of P. engelmannii in the region by providing in general, slightly higher correlation values for Tmax than for Tmean. Interestingly, the pattern observed in North America contrasts to many other temperature-limited regions of the Northern Hemisphere, where conifers have generally been noted to correlate more strongly to Tmean than to Tmax (observation made by the author team, results not published). Whether this is actually grounded in a tree physiological mechanism is still an open question. Furthermore, the general pattern revealed by the climate response analysis shows that the various dendroanatomical traits respond to consecutive temporal windows within a short seasonal window extending from June to August, in line with our understanding of the successive physiological processes (i.e., cell expansion and cell wall thickening) behind wood formation and growth (e.g., Fonti et al., 2013). These results support the climate-response pattern that has generally been observed for conifers across the Canadian Rockies (Luckman and Wilson, 2005) and the adjacent Interior British Columbia (Wilson and Luckman, 2003; Wilson et al., 2014), yet contrasts to the seasonally wide temperature imprint (extending between May and August and occasionally even between April and September) within latewood density of Picea mariana in the eastern Canadian taiga (Wang et al.,  2020). This is also the case when comparing our results with the previous study of Björklund et al. (2020) on latewood anatomical traits of Pinus sylvestris in northern Scandinavia, where the temperature response window extends from April to September. The narrow window of response patterns seen here is most likely constrained by the distinct and short warm season characterizing the climatology of the study site, where average monthly temperatures rise above 0 ∘C only in 5 months of the year (Fig. 1c).

We found that latewood-related parameters in general display a late summer (July–August) temperature sensitivity, while ring width and earlywood-related density and anatomy parameters most strongly correlate with mid-summer (June–July) temperatures (Fig. 4). The strongest temperature signals are found in anatomical components of the latewood, which are also the parameters displaying the highest RBAR statistics (Table 1). In particular, aMXD and Max. radial CWT stand out. The imprints of year-to-year temperature variability within these two parameters are, over the 1901–1994 period, very similar, if not identical, to those of the MXD derived from the X-ray technique. By comparison, the exceptionally weak inter-series signal strength of the MXBI parameter (Table 1) is compensated by high replication (N=182), and thus MXBI is also rather similar to aMXD, Max. radial CWT and X-ray MXD. However, the temperature signal of MXBI is shifted earlier by expressing a stronger correlation with July temperatures but weaker with August compared to aMXD, Max. radial CWT and X-ray MXD. The aggregated July–August temperature response of MXBI is thus in fact only marginally weaker than that of X-ray MXD, aMXD and Max. radial CWT.

Focusing only on anatomical traits with the highest temperature sensitivity (aMXD and Max. radial CWT), comparison with daily temperatures (Fig. 5) confirms a significant and strong mid-to-late summer signal over the 1880–1994 period. Breaking down the climate response in daily increments reveals that the strongest signal (r>0.5) occurs on average between days 192 and 251 of the year (i.e., 11 July until 8–9 September, with a peak correlation of 0.74 for Max. radial CWT and aMXD occurring between 21 July–20 August and 23 July–22 August), respectively. The temperature associations at the margins of the target season are, however, more unstable. We note, for example, that the September signal disappears around the first half of the 20th century for both anatomical parameters. A similar correlation structure holds for X-ray-derived MXD and to a lesser degree MXBI (N=182), but the two parameters exhibit enhanced correlation coefficients in the second half of the 20th century compared to the early period (also corroborated by the split-period calibration in Fig. 6). Moreover, despite the high sample replication, MXBI shows slightly weaker correlations with daily data than the other density-related parameters, particularly in the early 1880–1930 period, when ring widths coincidentally are the narrowest in the record (see Fig. 8 and Sect. 3.4). For comparative purposes we also include anatomically derived ring width, which shows on average the strongest correlations (r= 0.3 to 0.5) with temperatures between days 146 and 206 of the year (i.e., 26 May to 25 July).

The stability of the July–August temperature signals of detrended aMXD and Max. radial CWT, along with X-ray MXD and MXBI, were further assessed by a split-period calibration procedure (1901–1948 and 1949–1994) (Fig. 6). The two wood anatomical parameters calibrate more strongly to the early period compared to the late, when using both Tmean and Tmax. However, especially for Max. radial CWT, the calibration differences in the two periods are slight (r2= 53 % and 47 % against Tmax, respectively). By comparison, the X-ray MXD calibrates more strongly in the latter half of the instrumental period and shows more pronounced temporal instabilities (r2= 34 % and 55 % against Tmax, respectively). This contrasts to the prior finding (Luckman and Wilson, 2005) where no such instabilities in the early 20th century were detected. These contrasting results are most likely not related to using different climate data products because similar results (Fig. S3) were obtained when using the Luckman and Wilson (2005) temperature data, originally produced by the Meteorological Service of Canada. Instead we suspect that the discrepancy can be attributed to either using a larger network of MXD data than used in this study, or that Luckman and Wilson (2005) used multivariate regression models (including ring width and lagged growth responses) to explain a wider target season than attempted here.

Calibration trials with detrended data over the full period 1901–1994 reveal that Max. radial CWT performs overall best (Tmaxr2= 49 %), closely followed by aMXD (r2= 0.46 %) and X-ray MXD (r2= 0.46 %). The temporal instability of X-ray MXD and by comparison the robust and strong signals of the aMXD and especially the Max. radial CWT parameters are further confirmed by the resampling calibration trials presented in Fig. 6c, where 10 random series are drawn from the sample cohorts 1000 times without replacement, and the resulting parameter chronologies are subsequently correlated against July–August Tmax. The reason for the X-ray MXD loss in signal is difficult to disentangle, but it is unlikely related to having different samples for the X-ray and anatomical datasets because the resampling scheme clearly show that the r2 distributions are different (Fig. 6c) (also corroborated by a two-sample t test at a significance threshold of 0.05, indicating that the r2 statistics come from two populations with unequal means).

As shown in the previous sections, the climate imprint within the anatomical LW density components slightly differs from its X-ray and BI-based counterparts, although all these parameters essentially measure the same component in wood. As previously noted by Björklund et al. (2020), the main difference between these metrics is the measurement resolution, while factors such as the cell wall density is of marginal importance (Björklund et al., 2021). Thus, as part of a multiparameter approach, the higher resolution of dendroanatomy may serve to evaluate the potential risk of a resolution bias (in X-ray MXD and MXBI) when implementing these parameters both on shorter and longer time scales.

We have seen that the monthly correlations of the full MXBI dataset (N=182) differ slightly from the more physically direct density and anatomy parameters, which we hypothesize could partially be related to the lower measurement resolution that artificially makes it more similar to ring width and latewood width (Björklund et al., 2019). The pairwise correlation between parameter chronologies (Fig. 3) and the PCA biplot based on the percentile MXBI chronologies (Fig. S5) confirms this enhanced relationship with ring width and latewood width. To test this theory further, we have correlated the percentile MXBI chronologies against the target July–August Tmax (Fig. 7a) and against the full (N=182) detrended original ring width chronology (Fig. 7b), using resampling of data. Unfortunately, corresponding latewood width measurements are not available for MXBI, so this comparative analysis is restricted to ring width. Nevertheless, we find that when using the full July–August season the poorest temperature imprint is found in the MXBI values of the narrowest (∼40 %), and the widest (∼40 %) of the rings, while the strongest July–August signal can be recovered from the MXBI values in rings that are close to average in width (40th–70th percentile). Expanding the climate correlation analysis to monthly Tmax data (Fig. 7c) reveals, however, a gradual transition from a predominantly August temperature signal in the wide ring MXBI chronologies to a more July-dominated signal in the narrow ring MXBI chronologies. The MXBI values in rings that are close to average in width correlate equally strongly to both July and August, which explains the overall better performance of these data when comparing to the July–August target (Fig. 7c). Importantly, we find no correlation between the MXBI and ring width in the widest rings (Fig. S5). However, as we move towards narrower rings, the MXBI values become successively more like ring width/latewood width (Figs. 7b and S5). All in all, these results suggest that an effect of low measurement resolution may be present for narrower ring widths/latewood widths. If so, this means that the MXBI parameter may become subject to greater target season uncertainty, which may fluctuate between July and August signals over time, largely depending on the absolute ring width/latewood width of the analyzed tree ring sample collection and the resolution of the captured image. Although posing a challenge for paleoclimate reconstructions, this resolution issue is likely to become a less relevant methodological problem in a near future, as more laboratories are currently investing in the development of high-resolution image capturing systems and other analytical techniques to enhance the precision of the BI data.

Furthermore, we note that the correlations between the various latewood parameters against ring widths change from the early to late 20th century periods, and that the correlations slightly differ in magnitude and sign (Fig. 8). The MXBI is positively correlated with ring width, whereas the correlations for X-ray MXD range from non-significant to weakly positive. The Max. radial CWT, on the other hand, show a non-significant or weak negative correlation with ring width during the 20th century. This gradual, and slightly larger shift in moving window correlation against ring width during the early 20th century may thus be an indication that both MXBI and to some degree X-ray MXD are challenged by comparatively low measurement resolution. If this is the case, then the inter-annual climate signal may potentially become muted when ring (latewood) widths are narrow. This dependence could, in fact, affect the lower frequencies, and inflate multidecadal variability (Esper et al., 2015). Moreover, the fidelity to the monthly temperature targets may exhibit instability when rings (latewood) are narrow, shifting back and forth between August- or July-dominated signals (as seen in Fig. 7c). At the moment it is unclear how this phenomenon could affect the lower frequencies of our chronologies, as a robust picture of long-term trends in dendroanatomical parameters can only emerge from analysis of millennial length, multigeneration, composite chronologies suitable for RCS type detrending (Briffa and Melvin, 2011). Moreover, periods with persistence in narrow ring widths will force MXBI, and perhaps also X-ray MXD, to exhibit persistently low densitometric values (Björklund et al., 2019). Exacerbating this issue is that persistently narrow ring width and latewood width may not even be a product of the distinct and earlier temperature target (June–July, Fig. 4), but could also be related to stand dynamics/disturbances (Rydval et al., 2018), and thus pass down non-climatic distortions of decadal to centennial variations to X-ray MXD and MXBI. This clearly needs further scrutiny because it may be important for the interpretation of inferred climate signals back in time, particularly because the ring width correlation converges for the X-ray and anatomy data but dramatically diverges for MXBI (Fig. 8). The lower late period (1949–1994) signal of the anatomical parameters compared to X-ray MXD requires a different explanation (Fig. 6). According to the distribution of the r2 values in the resampling scheme of Fig. 6c, the late period Tmax signals are not appreciably different, so perhaps this is simply by chance compounded by having five times higher X-ray MXD replication.