Time to Use Dendrohydrological Data in Water Resources Management?

First, there is a mismatch in the temporal resolution of water resources and streamflow reconstruction models. The former work with daily to monthly decision-making time steps, while the latter are generally annual because they are constrained by the available tree-ring chronologies and their correlation with discharge data (e.g., Ho et al. 2017). In addition, streamflow reconstructions often focus on a specific season, rather than the entire hydrological year. For example, reconstructions from ring width typically target the growth season (e.g., D’Arrigo et al. 2011), yielding the cumulative discharge during the summer months. Both temporal resolution and continuity of the dendrohydrological data can therefore prevent their direct application in the quantitative decision-making models that dominate the water systems analysis literature.

Reconstructions often fail to capture extremes on the upper tail of the distribution of annual and seasonal flows. In warm regions, for instance, data derived from moisture-limited trees may not reflect saturated overland flow, which is necessary to reconstruct peak discharge events (Nguyen et al. 2020b). Another issue is that reconstructions over large spatial domains may fail to preserve the physical consistency of discharge in a river network. Suppose, for example, that we are modeling streamflow at two tributaries as well as the main stem of a river: the total flow of the tributaries should equal the flow on the main stem. This simple mass balance check is not explicitly accounted for by the vast majority of the reconstruction frameworks, generally based on point-by-point regression (Nguyen et al. 2020b)—where one models the discharge data independently for each station and relies on the proxy network to account for the spatial patterns. Because accuracy is problematic, streamflow reconstructions may generally be considered speculative, and therefore discounted in modeling and decision-making contexts.

Reconstructions are limited to pristine catchments or to areas for which naturalized streamflow data are available. The reason for this limitation is that streamflow reconstructions build on the numerical correlation between proxies and discharge data, so they are misled by anthropogenic interventions that alter the natural regime of rivers (e.g., hydropower operations, land use change, or irrigation water withdrawals). Because naturalized streamflow data are not widely available, many river segments across the world are precluded from reconstruction exercises. Paradoxically, it is in these regions with intense water consumption that dendrohydrological data could be most useful.

Finally, students in water resources engineering are not exposed to the rigorous analysis that determines the statistical features of tree-ring chronologies. Detrending and prewhitening, for example, are regularly used to remove multiple competing signals and temporal autocorrelation in raw tree-ring sequences, so that chronologies can be interpreted as evidence of hydrometeorological forcings (Fritts 1976; Cook 1985). But if we do not know which patterns have been deliberately preserved or excluded from the final product, we are liable to misinterpret the relation between chronologies and discharge data. This risk is well exemplified by Coulthard et al. (2020), who show how different detrending methods applied to the same ring width measurements can yield diametrically opposed chronologies, one emphasizing year-to-year variations in tree growth and one emphasizing century-scale variations in climate. An engineer working on the second chronology may erroneously conclude that the relation between annual discharge data and tree growth is weak.

A key prerequisite for a water resources management model is the availability of daily to monthly streamflow time series that capture, in a reliable manner, both pluvials and droughts. A fundamental advance in this area could build on the fact that different proxies have different sensitivities to climatic variability. For example, oxygen isotopes often correlate well with peak flow (Xu et al. 2019), while ring widths tend to capture the response to droughts (Gallant and Gergis 2011). One could therefore leverage the correlation between multiple proxies and different target variables to reconstruct subannual flows—a concept recently demonstrated by Rao (2020) and Nguyen et al. (2020a). Because more sampling sites and proxies are being developed (e.g., blue intensity, latewood density), we have an opportunity for creating models that bank on large multiproxy networks to reconstruct subannual streamflow. There are at least two research directions. First, there is a need to extend the spectrum of regression techniques so as to better capture the complex interaction between multiple proxies and discharge data. Linear regression frameworks may remain a cornerstone, but we can easily envision a future in which the modeling toolbox will include nonlinear regression techniques, such as Bayesian regression or state-space models (Rao et al. 2018; Nguyen and Galelli 2018). The next direction would be developing statistical disaggregation techniques that improve the temporal resolution of streamflow reconstruction. Ideally, the disaggregation process should be informed by a proxy network, as recently demonstrated in the monthly reconstructions developed by Stagge et al. (2018).

Future dendrohydrological models may also build on a completely different approach: one could reconstruct time series of precipitation and temperature—instead of streamflow—and use them to force a process-based hydrological model, as shown by Saito et al. (2015), Tozer et al. (2018), and Meko et al. (2020). The approach requires a large amount of data for the model calibration, but it has two advantages. First, process-based models implicitly account for the physical consistency of the reconstructed discharge data. Second, they can represent alterations in the water cycle due to operations of hydraulic infrastructure or land use changes, enabling reconstructions for nonpristine catchments. This approach could be leveraged with state-of-the-art large-scale hydrological and water resource models enhanced with realistic water management representation.

A last important area is the relation between dendrohydrology, water management, and the concept of nonstationarity. Dendrohydrological models build on the assumption that the relationship between proxies and discharge observed in instrumental records holds true over the entire length of the reconstruction. Such stationarity assumption has long been compromised not only by hydraulic infrastructure and land use changes, but also by anthropogenic climate change. In light of this expanded uncertainty, our community has progressively dropped the concept of the static design paradigm (Brown 2010) and developed robust and dynamic planning techniques that adapt to nonstationary trends (Herman et al. 2020). So, shall we still use dendrohydrological data to plan and operate hydraulic infrastructure in a nonstationary environment? In our opinion, the answer is yes. Considering the case of anthropogenic climate change, we have at least two technical challenges. First, we should not limit ourselves to compare the envelopes of reconstructed, observed, and projected streamflow variability. We need data analytics that quantify and reconcile the information contained in multiple power spectra and distributions. With such information, we could put future hydroclimatological risks into a better perspective, or develop a new generation of stochastic streamflow generators. Second, we should use the information provided by streamflow reconstructions to understand and, where possible, reduce the uncertainty in hydrologic projections informed by general circulation models (GCMs). A clear example is offered by the dynamics of the Asian monsoon, which are well captured by climate proxies (Goodkin et al. 2019) but poorly represented by the latest generation of GCMs (Aadhar and Mishra 2020). Streamflow reconstructions could therefore help us identify cases in which GCMs are not fully reliable, narrowing a major source of uncertainty for water resources management applications, such as flood risk management (Ziegler et al. 2020).