Decadal trends in seasonal streamflow and stream temperature across the United States
Abstract. Streamflow influences water supply and hydrologic extremes, whereas stream temperature controls thermal conditions critical for aquatic species. Changes in streamflow and stream temperature have far reaching ecological and socioeconomical implications; thus, it is essential to understand their interconnected seasonal trends and relation to external forcing factors. In this analysis, we (1) documented and compared decadal trends (1980–2020, 1990–2020, and 2000–2020) in seasonal and annual streamflow and stream temperature metrics at sites across the conterminous United States (CONUS), Alaska, Hawaii, and Puerto Rico; (2) assessed land cover, precipitation, air temperature, and baseflow fraction effects on streamflow and stream temperature; and (3) selected three hydrologic regions (Midwest, California-Nevada, and Atlantic Coast) within CONUS as regional case studies to better understand the effects of climate and land-use drivers on streamflow and stream temperature. Trends were evaluated using the following likelihood categories: likely (pt ≤ 0.1), somewhat likely (0.33 ≥ pt > 0.1), uncertain (0.67 ≥ pt > 0.33), and unlikely (0.67 > pt), where pt is the p-value determined from the Mann-Kendall trend test. Most trends were uncertain or unlikely, but some significant changes were found, such as positive correlations between air temperature and rising stream temperatures and between precipitation and increasing streamflow. Streamflow increased across most of CONUS for the 2000–2020 period, but decreases were more prominent for the 1980–2020 and 1990–2020 trend periods across the southern and western CONUS. During the 2000–2020 trend period, seasonal trends in increasing stream temperature were attributed to warmer summer and fall air temperature. Increasing seasonal streamflow was attributed to increases in precipitation in the winter and spring and decreases in baseflow fraction during the spring and summer seasons. We show streamflow trends vary in direction across different trend periods for multiple U.S. regions, which poses challenges for water resources planning and management. Regional results indicated that increasing streamflow in the Midwest region was attributed to increasing precipitation and decreasing baseflow fraction across all seasons. Decreasing streamflows in the California-Nevada region were associated with increasing spring baseflow fraction and increasing annual air temperature. Increasing and decreasing stream temperatures in the Atlantic Coast region were attributed to both increasing and decreasing air temperature, respectively. This study emphasizes the importance of studying streamflow and stream temperature together and demonstrates how comparisons of seasonal and annual metrics provide a more nuanced analysis of streams than annual metrics alone.
This study jointly examines trends in streamflow and stream temperature across the United States. The topic is interesting, and considering the two variables together at this scale is novel. The analysis is large-scale, the discussion is thorough, and the manuscript is generally well written and thoughtful. At the same time, I have substantial concerns regarding several methodological choices, the comparison among trend periods and, particularly, the strength with which some statistical associations are interpreted as attribution evidence.
Major comments
Although the manuscript reports Theil–Sen slopes, the presentation and synthesis, particularly in Figures 1–2, rely heavily on the fractions of sites assigned to p-value-based likelihood categories. In the maps, the sign of the slope determines the direction of change, but its magnitude is not displayed. From a hydrological perspective, trend magnitude and direction are at least as important as the likelihood classification. The same trend magnitude can receive a different classification depending on record length, variability, missing observations and the assumed stochastic model. Changes in the fractions assigned to these categories therefore do not necessarily represent corresponding changes in the underlying hydrological behaviour.
This is particularly important because the three periods (1980–2020, 1990–2020 and 2000–2020) are nested, have different lengths and share the same endpoint. Changing the starting year simultaneously excludes earlier decades, changes the estimated trend and its uncertainty, affects statistical power, and alters station coverage. The latter is especially relevant for stream temperature, for which the annual mean analysis is based on 26, 68 and 244 sites, respectively. Differences in the fractions of sites classified as “likely,” “uncertain” or “unlikely” therefore mix the effects of the selected time window, statistical power and monitoring network. The comparison is informative as an assessment of sensitivity to the starting year, but it should not alone be interpreted as evidence of temporal evolution or a recent shift. The manuscript should distinguish more clearly between changes in the estimated slopes and changes in their likelihood classification.
Figure 2 appears to show a recent tendency towards increasing streamflow across a large part of the United States, whereas this signal becomes less clear over the longer windows. This is an interesting result and may indicate that recent wetting is superimposed on longer-term variability or opposing behaviour in earlier decades. However, it requires a more nuanced interpretation than a comparison of significance categories alone. The use of all available stations is reasonable for describing each window, but a supplementary common-station analysis, where feasible, would help assess the influence of the changing monitoring network. If the objective is specifically to identify changes between periods, equal-length windows based on a common station set, or a formal change-point analysis, would provide a clearer basis for comparison.
P-values depend on the stochastic model adopted under the null hypothesis. They may change considerably when temporal dependence is included because positive autocorrelation reduces the effective sample size. Lines 149–153 indicate that a lag-one ARIMA comparison was used to select between an independent and an autocorrelated null model. However, the description is not sufficient to understand exactly how this choice modified the Mann–Kendall statistic and its resulting p-values. Please specify the detrending approach, the form of the autocorrelation correction, and how the selected null model was implemented.
Section 4.5 appropriately acknowledges that differences among trend periods may reflect long-term persistence and low-frequency natural variability. However, this point should be connected more directly to the statistical inference. If streamflow exhibits long-term persistence, extended natural fluctuations and apparent monotonic trends become more probable under a no-trend model than under independent or short-memory assumptions such as AR(1). Where such persistence is present, p-values based only on a lag-one or short-memory model may overstate the evidence for a trend. This limitation should also qualify the interpretation of short-period trends as persistent change.
The authors should reconsider the strength of some of their attribution claims. Linking precipitation to streamflow is hydrologically meaningful because the underlying mechanism is well understood; however, some of the other relationships are less directly supported. The treatment of baseflow fraction requires particular reconsideration. Baseflow fraction is calculated from baseflow and total streamflow and is therefore mathematically linked to the response variable. In addition, the baseflow estimates are obtained through a separation model applied to the streamflow observations themselves. It therefore seems problematic to treat baseflow fraction as an independent driver of streamflow. The discussion already offers a physically meaningful interpretation of these relationships in terms of changing flow partitioning and dependence on groundwater. However, this interpretation is not fully consistent with treating baseflow fraction as a driver elsewhere in the manuscript. The recurrent relationship between increasing streamflow and decreasing baseflow fraction may indicate an increase in direct runoff relative to baseflow, rather than an effect of baseflow fraction on streamflow. If feasible, the authors could examine baseflow and direct-runoff volumes separately; otherwise, these results should be framed as changes in flow composition rather than attribution to an independent driver.
Because the land-cover fractions sum to 100%, their trends are compositionally dependent: conversion from one class necessarily produces an opposing change in one or more other classes. This does not invalidate the observed associations, but it means that they should be interpreted jointly as land-cover transitions rather than as several independent pieces of evidence for separate drivers.
Minor comments
Figures 1 and 2: Please ensure that the colour coding of the trend categories is consistent between the maps and their legends.
Table 2: The inequality given for the “Unlikely” category appears to be reversed. According to the definitions in the text, this should be (pt>0.67). The same issue appears in several figure captions.
The term “Stable” in Figures 3–6 combines the “Uncertain” and “Unlikely” categories. I do not think that this is an appropriate interpretation. Failure to detect a monotonic trend is not evidence that the underlying process is stable. It would be preferable to retain the original category names or use a neutral expression such as “no supported monotonic trend.”
Figure 3: The combined graphical scheme remains difficult to follow. A short explanation in the main text or a clearer legend would help readers understand what is meant by “parameter increase,” “parameter decrease” and “parameter stable.”
Figure 6: The caption mentions the stream-temperature Theil–Sen slope, although the figure concerns streamflow. Please correct this and check the corresponding appendix captions for similar wording.