1 Introduction

Drought is a common feature of the climate system across the United States (US), impacting both human societies and ecosystems (Wilhite et al. 2007; Allen et al. 2010). For example, the 2012 drought was the most recent example of severe to moderate drought occurring in the Midwest United States (US; Hoerling et al. 2014). While the 2012 drought was severe in the Great Plains (Hoerling et al. 2014) and had substantial economic and ecological implications in the Midwest (Mallya et al. 2013; Roman et al. 2015), the intensity and duration of the drought was far less than previous instrumentally recorded events in the Midwest such as the extreme droughts of the 1930s and 1950s. Currently, a broad area of the Midwest is experiencing an ongoing pluvial that is in one of the wettest periods on record (Mishra and Cherkauer 2010; Ford 2014).

Tree rings have long been considered a robust proxy record of soil moisture variations throughout the US and other regions of the world (including humid areas such as the eastern deciduous forest biome), providing historical context to instrumentally recorded droughts (e.g. Cook et al. 1999, 2010; Maxwell et al. 2011; Pederson et al. 2013, 2014). Previously, reconstructed droughts in the region have been shown to be more intense and longer lasting than the more recent instrumentally recorded droughts (Cook et al. 1999, 2010). Previous studies highlight the need to assess the temporal stability of the climate-tree growth relationship and ensure climate signal consistency through time, most notably the diverging relationship between temperature and tree growth at high latitude sites (Jacoby and D’Arrigo 1995; Barber et al. 2000; Jacoby et al. 2000; Lloyd and Fastie 2002; D’Arrigo et al. 2004; Wilmking et al. 2004; Driscoll et al. 2005; Lloyd and Bunn 2007; Pisaric et al. 2007; Esper and Frank 2009). The “divergence problem” is not widespread and is largely an artifact at regional to local scales (Büntgen et al. 2008).

Changes in the drought signal recorded by tree-rings have been documented in a few studies and examining the stability in the relationship is becoming more important with ongoing climate change. However, past studies have either documented a shift in the seasonality in climate-growth response (De Grandpré et al. 2011) or weakening climate sensitivity in response to an infestation (Saladyga and Maxwell 2015). In this study, we provide the first documentation of an apparent deteriorating relationship between radial tree growth and soil moisture across an area of the Midwest US that is not accompanied by an increase in signal during another season. The declining relationship, referred to here as the fading drought signal, appears to be consistent across multiple species and sites within the region.

2 Materials and methods

2.1 Study sites and tree-ring data

During the period 2012–2014, we collected tree-ring data from five sites across the Midwest US (Fig. S1 and see Supplemental Material), including: three old-growth sites—Pioneer Mothers (PM), Donaldson Woods (DW), and Giant City (GC); one selectively logged site—Hoot Woods (HW); and one second-generation forest—Lilly-Dickey Woods (LD). From these sites, 15 chronologies were developed from tree species commonly found in the eastern deciduous forest (see Supplemental Material). All species (Table 1) are known to be sensitive to climate and their use in climate reconstructions is well demonstrated (Cook et al. 1999; Maxwell et al. 2015; Maxwell 2016). We used methods standard to dendrochronology for sample collection (see Supplemental Material).

Table 1 Tree ring statistics and validation statistics when using the stratified and randomized validation approaches for each species at the five sites
Full size table

2.2 Sample analyses

Tree cores were prepared using standard dendrochronological techniques (see Supplemental Materials). All chronologies were detrended interactively using signal-free standardization (Melvin and Briffa 2008) with a base curve being a “two-thirds” smoothing spline (i.e., it has an amplitude response of 0.5 at periods associated with 67 % of each series length; Cook and Peters 1981) to ensure that the climate variability signal was retained while the non-climatic variability was removed. Signal-free standardization was chosen after careful consideration of standardization techniques because it reduces “trend distortion” problems, where medium-frequency variance (i.e., decades to a century) can leverage the function used to remove non-climatic variance near the ends of the record (Melvin and Briffa 2008).

2.3 Climate data

We obtained climate data for the climate divisions that contained a study site (climate divisions Indiana 4, 5, and Illinois 8) from the National Climatic Data Center (NCDC) for the period 1895–2013, including monthly mean, mean minimum, and mean maximum temperature, precipitation, and the Palmer Drought Severity Index (PDSI; Palmer 1965). The PDSI is frequently used as the basis for dendroclimatic drought reconstructions (Cook et al. 1999; Cook and Krusic 2004; Pederson et al. 2012, 2013). To ensure the deteriorating signal was not an artifact of PDSI or its method of calculation, we also obtained gridded self-calibrated PDSI (scPDSI; Dai 2013) and gridded 6-month Standardized Precipitation Evapotranspiration Index (SPEI; Vicente-Serrano et al. 2010) for each grid that contained a site and then averaged the selected gridded data.

2.4 Climate analysis

We used the program DendroCLIM2002 (Biondi and Waikul 2004) to examine the possible changes in the climate response through time for each of the species chronologies. The program calculates correlation coefficients between each species chronology and the monthly NCDC climate division variables (temperature, precipitation, and PDSI) using 1000 bootstrapped samples. First, we conducted a 42-year moving window correlation analysis of the chronologies with PDSI for each month from April of the previous growing season to December of the current year. The deteriorating relationship is present with different window lengths (e.g., 75 years). We chose 42 years as a suitably long window and also because that is double the number of monthly predictors, which is the smallest length to allow enough degrees of freedom for the moving analysis (Biondi and Waikul 2004). The correlation values are reported for the last year of the window, creating a reporting period of 1936 to the last year of growth. Based on these results, we performed the analysis on the season with the strongest correlations (JJA) with tree-growth. Lastly, to further evaluate the JJA PDSI signal strength, we also calculated correlations of the tree-ring chronologies with monthly temperature and precipitation from the previous growing season April to current year December for the same moving window interval.

2.5 Impacts on climate reconstructions

To examine how changes in the relationship between tree growth and soil moisture impact climate reconstructions, we developed reconstruction models for each site and for the region. Tree-ring chronologies were used as the predictors to reconstruct summer (JJA) PDSI using a principal components (PC) regression analysis (Meko 1997; Cook et al. 1999 and see Supplemental Material). PC regression was used because it allows the development of multiple-species proxies, which have been shown to improve PDSI reconstructions in this area (Maxwell et al. 2015; note that GC has only one species—Q. alba).

Initially, model validation was performed using a split-period calibration-verification procedure that is common in dendroclimatic reconstructions (see Supplemental Material). For both late and early calibration models, we calculated validation statistics including R2 between the predicted and observed PDSI values for the verification periods, reduction-of-error (RE; Fritts 1976), and the coefficient-of-efficiency (CE; Nash and Sutcliffe 1970).

To determine how the time-varying drought signal influenced the site and regional climate reconstruction, we examined validation statistics for a series of iterative reconstructions by removing the most recent year and re-standardizing the chronologies at each iteration (i.e., 1895–2013, 1895–2012, etc.) using split-period calibration/verification. We conducted 35 iterations for each site reconstruction (with the exception of DW where the common period was shorter, thus 31 iterations), resulting in the last iterative reconstruction spanning from 1895–1979. Examining iterative reconstructions allows us to simulate validation statistics as if the tree-ring data were collected in 1979, 1980, or any year up to 2013. The period 1979–2013 is also the period of the fading signal, allowing us to determine the timing and impact of the deteriorating signal on model validation. Using the iterative procedure also helps to evaluate whether the deteriorating signal is an artifact of a particular detrending (our use of signal-free standardization also addresses this issue). If end-of-record distortion were the cause of the deteriorating signal, all iterations would have poor validation statistics, as opposed to those that only use the most recent data.

We also evaluate a randomization procedure for validating reconstructions. This approach is comparable to but distinct from ensemble approaches (Li et al. 2007; Wahl and Smerdon 2012) that allow for the confidence of the reconstruction to be evaluated. Ensemble approaches are useful because they use the behavior of the time series to generate error distributions. A limitation, however, is that (as implemented) they do not characterize how different portions of the “calibration-validation space” influence model performance (i.e., they consistently use the same data for calibration and validation). Our primary goal is to evaluate how different calibration-validation samples influence the relative performance of the model. We use a procedure that is similar to the traditional split-period method except that the samples (halves) are selected randomly and the calibration/validation procedure is repeated a large number of times (here, using 1000 random samples). From these random samples, we develop validation statistics and confidence intervals. This approach contextualizes the split-period validation statistics in two ways: (1) it avoids the possibility of selecting halves such that one or both are unrepresentative of conditions during the entire period of record and (2) it places the validation procedure within a probabilistic framework (similar to the ensemble approach). If the data have strong serial correlation, however, there is potential bias introduced by the two samples not being independent. As a result, we recommend that this method be used with the split-period calibration and that autocorrelation functions (acfs) be inspected. Importantly, the randomization procedure usually samples blocks or runs of data. As most tree-ring and climate data (especially in the eastern US) have acfs that are near-zero at lags of 2 years or longer (Fig. S2), any data values separated by more than 1 year are effectively independent. The acfs, therefore, portray a worst-case scenario as the random sampling approach would have to use every other year of the time series to have bias as large as that suggested by the lag-1 autocorrelation coefficients.

To examine the potential impact of the absence of extreme drought over an extended period of time, we also use a stratified-by-every-other-ranked-value approach to separate the PDSI and tree growth data into two representative samples. In this way, PDSI values that range from driest to wettest (and their corresponding tree ring values for each chronology) are included in each stratified sample. Data associated with the first sample of this ranked/stratified list of PDSI (i.e., the 1st, 3rd, 5th, etc. values) are selected and grouped (hereafter G1) and used to calibrate a linear regression model between PDSI and tree growth. We used the remaining half of the data (G2, which are associated with the 2nd, 4th, 6th, etc. PDSI values) to verify the model by using the G1-calibrated model to predict the values of G2. The same procedure is repeated using G2 to calibrate and G1 to verify the model. We then use the distribution of CE values derived from the randomized approach to provide a context for the CE values from both the traditional split-period method and the stratified approach. Importantly, both the randomized and stratified procedures test the ability of proxy-based models to interpolate missing values in a climatic time-series, whereas traditional split-period methods test the ability to extrapolate from one time period to another. We therefore caution that validation statistics from the randomized and stratified procedures are primarily used to provide additional context and may not be representative of a climate reconstruction.

3 Results and discussion

3.1 Moving correlation analysis

All core samples for each species had significant interseries correlation of their standardized ring widths (p < 0.001), indicating that individual trees from the same species were growing similarly and responding to environmental factors (Table 1). Time series of PDSI with tree growth showed that tree growth from co-occurring species tracks PDSI well until ≈1980 where tree growth and PDSI become decoupled (Fig. 1). Moving-window correlations showed strong relationships between all species chronologies and PDSI for May through September, with the strongest relationship existing during JJA (Fig. S3). The strength of the relationship between JJA PDSI and tree growth, however, weakened through time in all species sampled. Of the five locations sampled, DW and GC appear to have the largest decrease in the strength of the relationship through time, while LD and PM have less severe decreases (Fig. S3). The decrease in the strength of the relationship is relatively consistent and did not appear to be associated with a single species or site. While Q. alba at DW, LD, and PM faded less, its signal declined dramatically at GC. Similarly, L. tulipifera showed a weakening but still significant relationship through time at LD and HW, but its relationship declined much more at DW and PM. The decrease in the strength of the relationship was not matched by an increase in signal strength in the other months (Fig. S3). In addition, the deteriorating signal was not related to the climate-division version of PDSI, as both scPDSI and SPEI showed similar weakening (Fig. S4).

Fig. 1
figure 1

Timeseries of standardized tree-rings and PDSI (black line) for a) Donaldson Woods, b) Pioneer Mothers, c) Hoot Woods, d) Lilly-Dickey Woods, and e) Giant City. Each species is represented by the same line color across sites: Q. alba (blue), Q. velutina (red), Q. rubra (green), L. tulipifera (cyan), J. nigra (magenta), and Q. prinus (yellow)

Full size image

The age of the trees also did not seem to have an impact on how much the signal deteriorates based on the expressed population signal (EPS; Wigley et al. 1984) where EPS values of greater than 0.85 indicate that there is good sample depth (Table 1). One of the two sites that had the most stable relationship through time had older trees (PM, with trees >250 years) while the other was a second-generation forest with younger trees (LD with trees ≈120 years old).

The moving-window correlation analyses for precipitation and temperature also demonstrated that summer (JJA) months had the strongest relationship with tree growth (Figures S5 and S6). The relationship between precipitation and tree growth through time was inconsistent between species. While many species showed a similar declining signal of JJA precipitation as with summer PDSI, some (e.g., LD_LITU, LD_QUPR, HW_LITU, and PM_JUNI) exhibited an increase in the strength of the relationship and others (e.g., LD_QUAL, PM_QUAL, PM_LITU, and PM_QURU) demonstrated a consistent relationship through time (Figure S5).

The moving correlation for temperature revealed June to be the most important month for tree growth, followed by July and August (Figure S6). The strength of the relationship through time decreased consistently for all sampled sites and species, with GC_QUAL being the only chronology with a consistent relationship. In general, the summer temperature signal was not as strong as the PDSI signal and the weakening was not as temporally consistent between species in the change point; however, the temperature response seemed to be more influential than that of precipitation. To investigate the temperature relationship in more detail, we used averages of JJA maximum and minimum temperatures separately to perform moving correlations with standardized tree-growth. For all species, maximum JJA temperatures and growth initially had strong negative correlations ranging from −0.4 to −0.8 in earlier iterations and declining to 0.0 to −0.4 when more recent data are included (Fig. 2b). Minimum JJA temperatures were not as strongly correlated with growth; however, they went from generally negative to positive through time (Fig. 2a). Interestingly, temperature during JJA in this region has a slight downward trend through time, while JJA precipitation and PDSI had trended towards wetter conditions. An extended period of time where severe heat and drought conditions are not present could create conditions where tree growth is not as sensitive to soil moisture, particularly for some species in mesic deciduous forests (Brzostek et al. 2014).

Fig. 2
figure 2

Moving correlations of standardized radial growth with JJA minimum temperature (a) and maximum temperature (b) for all species chronologies. Dotted lines represent a non-significant relationship. The year represents the last year of each 42-year moving window

Full size image

Averaging PDSI values over multiple months, particularly JJA, provided a stronger relationship (Cook et al. 1999; Maxwell et al. 2015), so we used JJA PDSI to re-estimate the moving correlation with the species chronologies. Correlations with JJA PDSI showed a similar pattern of declining signal strength through time (Fig. 3). The sites with the most stable relationship through time across species were PM and LD, but even these sites exhibited a decrease in the correlation between tree rings and JJA PDSI from ≈0.65 to ≈0.4. Further, at least one species at PM and LD (L. tulipifera for PM and Q. velutina for LD) declined to the point of no longer being significant (Fig. 3c and e). HW did not have a monotonic decline, but nonetheless had a fading signal such that Q. rubra correlations decreased until the relationship was not significant in recent decades (Fig. 3d). DW showed the most consistent decline in correlation (from 0.7 to 0.2), with two species (Q. rubra and L. tulipfera) fading until the relationship was no longer significant (Fig. 3b). Of the species studied, L. tulipfera, Q. rubra, and Q. velutina exhibited the greatest declines while Q. alba showed more modest decreases across sites. However, Q. alba at GC showed a dramatic decrease in the PDSI signal – dropping from correlations of ≈0.7 in the early part of the record to less than 0.3 in the most recent windows.

Fig. 3
figure 3

Moving correlations of standardized radial growth with JJA PDSI for a) Giant City Q. alba, b) Donaldson Woods L. tulipifera, Q. alba, Q. rubra, and Quercus velutina, c) Pioneer Mothers L. tulipifera, Q. alba, Q. rubra, and J. nigra, d) Hoot Woods L. tulipifera, Q. rubra, and e) Lilly-Dickey L. tulipifera, Q. alba, Q. prinus, and Q. velutina. Dotted lines represent a non-significant relationship. The year represents the last year of a 42-year moving window

Full size image

3.2 Impact on climate reconstructions

The scatterplots showed that PDSI in the most recent 20-year period has been mostly positive (Fig. 4; inspection of instrumental precipitation and temperature data confirm that the most recent 20 years were relatively cool and wet). These results suggest that the more recent period is not only cooler and wetter than average but also has a nearly complete lack of extremely dry years (i.e., PDSI < −3), which may produce the fading signal for most species. In recent decades, just two droughts affected the region, both of which were modest in magnitude (i.e. PDSI < −2 during JJA of 1988 and 2012) and not comparable to the much more severe drought years of the 1930s and 1950s in terms of intensity or duration. The periods that contain the 1930s and 1950s drought events contain a much stronger relationship with soil moisture (Fig. 3).

Fig. 4
figure 4

Color-coded scatter plots showing the relationship between standardized species radial growth and PDSI for 20-year increments. The most recent 20-year segment is represented in yellow while the earliest is blue. All data are used to generate the linear regression line

Full size image

Our site and regional JJA PDSI chronologies consisting of multiple co-occurring species show that the fading drought signal has a deleterious impact on climate reconstructions (Fig. 5). Calibrating the iterative site and regional JJA PDSI reconstructions on the early half of the record shows that RE values are the lowest (but still positive; Fig. 5a) and CE values are lowest (Fig. 5b) when verification uses data from the period 1955–2013, with both increasing as more recent years are removed. HW CE values are the least robust through time, while DW and PM have acceptable CE values until they are barely positive starting in ca 2003 and 2010 respectively. However, CE values continue to decline to the point of being negative for HW and DW after 2008. LD has the strongest temporal stability in CE values, however they too trend weaker in the iterations (Fig. 5b). Similarly, the explained variance declines through time with the lowest R 2 values coming from the models with the most recent data (Fig. 5c). The site reconstruction models that only include data from before 1980 explain a similar and stable amount of the variance in JJA PDSI (≈55 %), similar to validation statistics found in previous drought reconstructions from this region in the early 1980s (e.g., Cook et al. 1999). However, when the iterative process includes the more recent data, all of the site reconstructions decline with the best site models (PM and LD) explaining approximately 30 % and the poorest sites declining to ≈26 % (HW) and ≈15 % (DW). When combining all the chronologies into a regional PDSI chronology, the iterative models show declining RE and CE values and a decrease in explained variance from ≈70 % to 38 %, indicating that increases in the number of chronologies improved the model but the decrease in the strength of the relationship is still present (Fig. 5c). Calibrating the iterative JJA PDSI reconstructions on the late period showed stable positive values of RE and CE through time (Figure S7A and S7B), with HW being consistently marginally positive due to the weak climate signal at that site. Further, the R2 values from the late calibration model did not decline (Figure S7C) with the sites consistently explaining approximately 58 % (DW), 56 % (PM), 45 % (LD), and 28 % (HW) of the variance in JJA PDSI. The regional JJA PDSI reconstruction showed similar stability in RE, CE and R 2 values explaining ≈60 % of the variance in PDSI. The models give better validation statistics when validated on the data that have extreme drought conditions.

Fig. 5
figure 5

Validation statistics from the iterative climate reconstructions using the early calibration period for RE (a), CE (b), and R2 (c) for the multiple co-occurring PDSI reconstructions for each site and regionally (“Regional”)

Full size image

Using the randomized validation approach rather than the split-period method produces a distribution of CE values allowing for comparisons of the stratified and traditional validation methods (Table 1; Fig. 6). When ensuring that severe drought is present in both periods of the calibration-verification models (i.e., the stratified approach), all the resulting CE values are positive and are located in the peak of the randomized CE distribution. The traditional method of validation results in CE values distributed sporadically—and, in some cases, unrepresentatively—along the distribution of randomized CE values (Fig. 6). While, part of the difference in values can be explained from the interpolation versus extrapolation methods of the two validation statistics, split-period (i.e., extrapolation) methods have given excellent validation statistics for PDSI reconstruction from this region in the past (Cook et al. 1999).

Fig. 6
figure 6

Histograms of randomly generated CE values for each species with 95 % confidence intervals marked in blue bars. Traditional split-period validation results are shown in cyan (early half calibration) and green (late half calibration). Validation results from the stratified approach are shown with red lines

Full size image

The lack of extreme drought conditions in the more recent period produces lower/negative RE and CE values when using the early period to calibrate the model for the split-period method. Extreme drought years have the most leverage in a regression model, thus when a long period is absent of extreme values the strength of the linear relationship declines as shown in the scatter plots (Fig. 4). An extended period of time absent of extreme droughts has created an environment where trees appear to be less sensitive to the relatively low variability in soil moisture.

Interestingly, Fig. 1 indicates that growth trends between species are fairly consistent until the period of recent weakening when tree growth varies dramatically between species. Thus, one possible explanation of the deteriorating relationship is the differing methods by which trees use water. Recent work has shown a difference in water-use strategies ranging from isohydric to anisohydric (Choat et al. 2012; Manzoni et al. 2013; Klein et al. 2014; Martínez-Vilalta et al. 2014). Brzostek et al. (2014) found that chronic mild water stress influenced growth declines in mesophytic species (e.g., L. tulipifera) more than in xerophytic species (Quercus spp). When conditions are extremely dry, growth from all species is negatively impacted. However, other anthropogenic factors also could be influencing this relationship, such as increases in carbon dioxide concentrations, land-use change, and nitrogen availability. More work is needed to investigate how these variables interact and to determine if the deteriorating signal is a temporary phenomenon.

4 Conclusions

We find that multiple species in the eastern deciduous forest from the Midwest US appear to have become less sensitive to soil moisture in recent decades. We attribute the apparent fading signal to a sustained period without extreme drought. Thus, trees in this region have not lost the ability to give historical context to modern drought in the Midwest. Yet, if the recent pluvial continues, the linear relationship between soil moisture and tree growth will likely continue to deteriorate, which would have serious implications for future dendroclimatological studies in this region. However, when extreme drought returns in this region we expect the growth sensitivity to increase as larger water stress more effectively and consistently restrains tree growth across species.

We highlight the need for more research to understand the spatial dynamics of the fading drought signal. Future research will concentrate on (1) determining whether this phenomenon is local to the Midwest region or occurs across a more geographically extensive area (i.e., eastern US); (2) examining other mid-latitude mesic forest regions that have a range of recent drought conditions; and (3) providing a better understanding of species-specific responses to drought in order to disentangle the causes behind the inconsistent radial growth response between species during the pluvial. With a better understanding of the fading drought issue in the Midwest, and potentially across the eastern US, the annually-resolved, precisely-dated, and well-replicated proxy of tree rings should continue to provide historical context to current and future hydroclimate conditions.