Climatic Controls on Mean and Extreme Streamﬂow Changes Across the Permafrost Region of Canada

: Climatic change is affecting streamﬂow regimes of the permafrost region, altering mean and extreme streamﬂow conditions. In this study, we analyzed historical trends in annual mean ﬂow (Q mean ), minimum ﬂow (Q min ), maximum ﬂow (Q max ) and Q max timing across 84 hydrometric stations in the permafrost region of Canada. Furthermore, we related streamﬂow trends with temperature and precipitation trends, and used a multiple linear regression (MLR) framework to evaluate climatic controls on streamﬂow components. The results revealed spatially varied trends across the region, with signiﬁcantly increasing (at 10% level) Q min for 43% of stations as the most prominent trend, and a relatively smaller number of stations with signiﬁcant Q mean , Q max and Q max timing trends. Temperatures over both the cold and warm seasons showed signiﬁcant warming for >70% of basin areas upstream of the hydrometric stations, while precipitation exhibited increases for >15% of the basins. Comparisons of the 1976 to 2005 basin-averaged climatological means of streamﬂow variables with precipitation and temperature revealed a positive correlation between Q mean and seasonal precipitation, and a negative correlation between Q mean and seasonal temperature. The basin-averaged streamﬂow, precipitation and temperature trends showed weak correlations that included a positive correlation between Q min and October to March precipitation trends, and negative correlations of Q max timing with October to March and April to September temperature trends. The MLR-based variable importance analysis revealed the dominant controls of precipitation on Q mean and Q max , and temperature on Q min . Overall, this study contributes towards an enhanced understanding of ongoing changes in streamﬂow regimes and their climatic controls across the Canadian permafrost region, which could be generalized for the broader pan-Arctic regions.


Introduction
The Arctic and subarctic regions of the world-mostly underlain by continuous/ discontinuous permafrost-are affected by a range of extreme streamflow conditions, which include low to no flows in winter under ice-cover; spring floods due to river-ice breakup and snowmelt; snowmelt driven peak flows during spring/early summer; and in rare instances, rainfall driven peak flows in late summer. These extreme conditions are influenced by various climatological and landscape drivers and controls, and changes in these factors affect the magnitude, timing and duration of such events. Particularly, enhanced warming in the northern latitudes (which is almost twice the global average temperature increase [1][2][3]) and amplified poleward moisture transport to the region [4,5] (which propagates into increased precipitation) are affecting different components of the regional hydro-climatic systems. Prominent changes include reductions in the magnitude, duration and extent of snow cover [6][7][8], enhanced permafrost thaw [9,10] and changes in the snowfall-rainfall balance [11], all with the potential to alter the mean and extreme streamflow conditions across the permafrost region.
Trends in streamflow, precipitation and temperature variables were analyzed using the R "zyp" package [38] by employing a non-parametric Mann-Kendall test [39] with iterative pre-whitening [40]. The iterative pre-whitening method, employed to remove the effects of serial correlation, was used because it has been found to be robust in handling Type-I error and produce conservative estimates of p-values, thus limiting false trend detection [41]. Besides accounting for serial correlations for each variable and at each individual station, the effect of spatial correlation among stations (field significance) was evaluated by using the false discovery rate technique suggested by Wilks [42]. Results of the analyses are presented for 10% (p < 0.10) and 5% (p < 0.05) significance levels, with the former considered as the level of statistically significant trend, while the latter signifies a stronger trend.
We analyzed the spatial relationships amongst climatological means of the temperature, precipitation and streamflow variables as well as their trends using the Kendall's τ correlation. This method was preferred to the more common Pearson correlation because Kendall's τ uses a rank-based procedure, is more resistant to outliers, and measures all monotonic (linear and nonlinear) correlations [43]. All streamflow values and trend values were normalized by their respective drainage areas when comparing them with precipitation and temperature values/trends.
It should be noted that, in most cases, the drainage areas of most hydrometric stations extend beyond a particular permafrost zone, and in some cases, to the permafrost-free zone. Therefore, analyses specific to different permafrost zones were not undertaken. However, we considered trends for seven (out of 11) of the Water Survey of Canada's drainage regions [37], which include: (3) Labrador, Northwestern Quebec; (4) Albany; (5) Nelson; (6) Arctic, Hudson Bay; (7) Upper Mackenzie; (8)(9) Pacific and Yukon; and (10) Arctic, Lower Mackenzie and Queen. Region 8 has only one station and its results were combined with region 9.
For selected stations, we characterized temporal sensitivities of streamflow variables to climatic variables using the MLR framework. Precipitation, temperature and their interaction were considered as the driving variables and streamflow components were considered as the response variables, given by: where β 1 , β 2 , β 3 and β 4 are the coefficients of MLR. ∆Q, ∆P and ∆T are the anomalies of streamflow, precipitation and temperature variables with respect to their mean values over the period 1976 to 2005, with ∆Q and ∆P considered as the percentage change and ∆T considered as the absolute change. ∆P∆T represents the interactions between precipitation and temperature. Lehner et al. [44] and Chegwidden et al. [45] used a similar concept in their evaluations using MLR models, although the latter did not use the interaction term. Further, we used the R "Caret" package [46] variable importance (VI) metric, which is based on the absolute t-statistic of each model coefficient, to assess the relative sensitivity of each input variable with respect to the output variable of the MLR model. Figure 1 depicts the trends in Q mean , Q min and Q max along with the Q max timing at 84 hydrometric stations across the permafrost region of Canada. The results generally corroborate with previous studies of six large Arctic flowing rivers (Appendix A Table A1), suggesting trends in the Canadian permafrost region are consistent with those across the pan-Arctic. Specifically, similar to the trends across the pan-Arctic, there are larger fractions of stations with increasing Q mean (62%) and Q min (68%) trends than decreases, although some of the trends are not significant. The larger fraction of stations (44%) with significant increases (p < 0.10) in Q min is also consistent with the previous studies. Overall, Q min increases represent the most prominent trend in this study, with 38% of trends field significant (Table 1); thus, most trends are globally significant. Most stations with increasing Q mean trends in this study are located in north-western Canada, while the southern and eastern regions show no trends, or even a few decreasing trends. Likewise, while there are few stations with significant trends in Q max and Q max timing in north-western Canada, the majority of stations in other regions show no significant trends. Furthermore, number of stations with field significant trends are even smaller, implying the lack of globally significant trends in these variables. This is especially the case for Q max timing, for which none of the trends are field significant, thus, these trends are practically insignificant. Nevertheless, the effect of increasing Q min is reflected in declining Q max to Q min ratios for some of the stations (Table 1). Specifically, 33 out of 79 stations exhibit significant downward trends, with field significant trends for 19 stations.

Spatial Variations of Streamflow Trends
Considering the trends for the seven drainage regions of Canada, it can be seen that the bulk of the stations are located in western Canada, with only a limited number of stations in eastern Canada ( Table 2). Given the limited number of stations, field significance testing was not done at the regional level. Considering the temporal trends, regions 8, 9 and 10 have larger fractions of stations with trends in Q min and Q max /Q min , while regions 3, 4, 5, 6 and 7 have smaller fractions of stations with trends. Some of the trends in region 3 are opposite to other regions, e.g., while there are declining Q mean and Q max trends in region 3, the declining trends are less prevalent in other regions. However, given the limited number of stations, generalization/extrapolation of patterns for these data-poor regions (regions 3 to 5) is considered problematic. There are also regional differences in western Canada; e.g., in the upper Mackenzie (region 7), there is a larger fraction of stations with increasing Q min trends compared to the lower Mackenzie (region 10).  Regions 9 and 10 consist of the Yukon and Mackenzie rivers, respectively, along with their tributaries. The streamflow trends at the most downstream stations and upstream stations are generally consistent. Specifically, the significantly increasing Q min trend in the Yukon Eagle station (1951 to 2018) is consistent with 9 out of 11 upstream stations. The Yukon Eagle station does not have significant trend in Q mean , Q max and Q max timing, which is in general agreement with the lack of trends in 8 (12), 10 (12) and 10 (12) upstream stations, respectively (numbers in the parentheses are the total number of stations). Similar results were obtained when comparing the 1973 to 2016 trends for the most downstream station on the Mackenzie River (Mackenzie River at Arctic Red River) with upstream stations. Specifically, the absence of significant trends in Q mean , Q max and Q max timing are in agreement with 14 (15), 14 (18) and 15 (18) upstream stations, respectively. However, the downstream station exhibits no significant trend in Q min , while 11 (15) upstream stations show increasing trends. It is also noteworthy that for the Mackenzie River streamflow variables, the lack of significant trends over 1973 to 2016 are in agreement with Yang et al. [47] over 1973 to 2011. For the Yukon Eagle station, significant increases in Q min and insignificant change in Q max for the period 1951 to 2018 agree with Bennett et al. [48], in which the 1954/1964 to 2013 periods were used. However, insignificant change in the Yukon Eagle Q mean for the period 1951 to 2018 differs from the significant increasing trend for the period 1975 to 2016 according to Box et al. [27].
Given that the Q min increase is the most prominent trend in this study, as well as other studies across the pan-Arctic (see Appendix A), an important question arises on its link to permafrost degradation. In this respect, while a majority of stations in the discontinuous permafrost zones in western Canada show increasing Q min trends, only a few stations in the discontinuous permafrost zones in Saskatchewan, Manitoba and Ontario show increasing Q min ( Figure 1). Furthermore, Smith et al. [21] found greater Q min increases in the non-permafrost zone than permafrost zone of Eurasia and speculated that decreased seasonal freezing of soils caused by warmer winters and/or deeper snowpack may be causing the Q min increases. Such processes could also be contributing to Q min increases in this study, and consequently, the decline of the Q max -to-Q min ratio.  Table 2). Specifically, while the stations with temperature increases are distributed across all regions, the bulk of stations with increasing precipitation trends are located in region 10 followed by region 7. Some stations in regions 5 and 6 also show an increasing trend in O-M_P.

Spatial Streamflow and Climatic Trends Relationships
Next, we present correlations among the 1976 to 2005 climatological means of streamflow components and driving precipitation and temperature ( Figure 3a), with Q mean , Q min and Q max values normalized by the basin areas (hereafter referred to as the q mean , q min and q max ). The correlations specify the strength of spatial relationships, such as the significant positive correlations among q mean , q min and q max suggest that the basins with higher values in one flow component are also likely to have higher values in the other two components. Furthermore, as precipitation is the main driver of annual flow, q mean for the basins are positively correlated with both O-M_P and A-S_P. The negative correlation of q mean with A-S_T is likely due to higher evapotranspiration losses in warmer basins. q max is negatively correlated with its timing, which is probably due to the wet-to-dry precipitation gradient, and warm-to-cold temperature gradient from south to north in the permafrost region [49]. Thus, the warmer southern basins are likely to produce higher q max , and earlier timing than the colder northern basins. On the other hand, the negative correlation between q min and q max could be linked to snowpack and baseflow generation. For instance, the relatively colder basins are likely to have lower baseflow and larger snowpack than warmer basins with a similar precipitation amount, thus producing relatively smaller q min and larger q max in colder basins. Besides the spatial variability of precipitation and temperature, the seasonality also plays a role on streamflow responses. Specifically, the positive correlations of q min and q max with O-M_P are in line with the expectation that the basins with higher winter precipitation also produce higher flow rates. q min also has a positive correlation with O-M_T. Therefore, the higher the winter temperature in the basin, the higher the low-flow response, likely due to higher rainfall-to-total-precipitation ratio and higher subsurface flow. Furthermore, the negative correlation of q max with A-S_T is likely related to spring snow accumulation and melt, which is constrained by spring temperature; thus, the higher the A-S_T, the earlier the snowmelt and the smaller the q max . However, although warmer basins are generally expected to have earlier Q max timings, there is no significant correlation between Q max timing and A-S_T.    higher flow rates. qmin also has a positive correlation with O-M_T. Therefore, the higher the winter temperature in the basin, the higher the low-flow response, likely due to higher rainfall-to-total-precipitation ratio and higher sub-surface flow. Furthermore, the negative correlation of qmax with A-S_T is likely related to spring snow accumulation and melt, which is constrained by spring temperature; thus, the higher the A-S_T, the earlier the snowmelt and the smaller the qmax. However, although warmer basins are generally expected to have earlier Qmax timings, there is no significant correlation between Qmax timing and A-S_T.   The correlations among streamflow and climate variable trend values (Sen's slope) signify the relationships in the magnitude and direction of changes (Figure 3b). The correlogram reveals significant but mostly weak correlations (Kendall's τ < 0.2) for some of the variables. The positive correlations amongst q mean , q min and q max trends imply that the increase in one of the streamflow variables is likely accompanied by increases in other variables, although the trends for individual stations may not always be significant. With reference to the driving variables, q max and q mean trends show weak positive correlations with O-M_T trends. Such unidirectional flow and temperature trends appear counterintuitive, as an increased temperature can normally be expected to lead to decreased snowpack and consequently decreased flows. In this case, other factors that accompanied increasing temperature trends may have played a role, such as precipitation trends, although the correlations are not significant. The positive correlation of q min with O-M_P trends indicate increasing low-flow with increasing winter precipitation. The significant negative correlations of Q max timing trends with A-S_T and O-M_T trends are in line with the expectation of earlier timing with increasing temperature, which cause smaller snowpack and earlier snowmelt. Furthermore, the negative correlations between q min and Q max timing trends are likely linked to temperature and snowpack changes, because increasing winter temperatures lead to larger q min and smaller snowpack, with earlier snowmelt leading to earlier Q max timing. However, the expected linkage between q min increases with temperature increases were not found to be significant.
Furthermore, a closer examination of the temperature, precipitation and streamflow changes reveals expected patterns. For instance, 31 of the 35 significantly increasing q min trends coincide with the O-M_T increasing trend. However, only a few q mean and q max results showed significant trends, and these trends do not always line up with the significant O-M_P and A-S_P trends. A number of factors affect the presence/absence of the relationships between streamflow and climate variable trends. Firstly, streamflow trends reflect the cumulative effects of precipitation, temperature and snow trends. Therefore, trends in one of the driving variables (temperature or precipitation) may not lead to trends in streamflow variables. Secondly, given the complex interactions between the climatic drivers, the trend values may not change monotonically for different variables, thus resulting in weak or no correlations with the streamflow trends.
There are also a number of sources of uncertainty that could lead to discrepancies and inconsistencies. A major source of uncertainty is the representativeness of the precipitation and temperature in the PNWMAmet and NRCanmet dataset, particularly due to the sparse station network in the high-latitude region [50]. Additionally, uncertainties arise from the quality of the observed discharge data, especially during the ice-covered low-flow period [51], as well as missing values in the discharge data. There are also inconsistencies in the length of the records. Given that some of the Q min , Q mean , Q max and Q max timing records have gaps, the compared lengths are not always consistent with temperature and precipitation records. Nevertheless, the analyses of climatological means and trends of the driving and resultant variables generally provide physically plausible explanations.

Climatic Controls on Streamflow Variables
We focused our temporal analyses on stations with the longest available flow records, which included: The sensitivities of precipitation and temperature changes on Q mean , Q min and Q max anomalies relative to 1976 to 2005 together with the results of MLR analyses are shown in Figure 4. Note that Q max timing results are not shown because of the overall lack of relationships with climatic drivers. The relative controls of precipitation and temperature on the three streamflow variables are depicted in terms of VI scores, with R 2 summarizing the MLR model fits. We also explored different combinations of the precipitation and temperature seasons in the MLR model setup. The results that yielded the best R 2 values were depicted, which included annual values (Ann_P and Ann_T) for Q mean , October to March values (O-M_P and O-M_T) for Q min , and October to July (O-J_P and O-J_T) values for Q max for both stations. These best combinations of months are in agreement with the analyses of climatic controls on these three streamflow variables using the historical and projected future temperature and precipitation for the Liard River basin [14]. The relative controls of temperature, precipitation and their interaction term are depicted by VI 1 , VI 2 and VI 3 , respectively.
The MLR model results for Q mean indicate the dominant controls of Ann_P (VI 2 > 0.75) for both basins (Figure 4a,b). While Ann_T has a small influence on Q mean for Liard-LC, its influence for Yukon-E is negligible. The controls of precipitation-temperature interactions on Q mean for the two basins are non-existent to negligible. Furthermore, as illustrated by the stratified patterns of Q mean anomalies relative to Ann_P anomalies, the sensitivities of Ann_P on Q mean are positive for both basins; i.e., the larger the Ann_P, the larger the Q mean . Overall, the model performance for Q mean is good for both basins (R 2 > 0.7), i.e., Ann_P and Ann_T anomalies can explain over 70% of variance in Q mean anomalies.  In the case of Q min , temperature plays a larger role, with VI 1 for O-M_T > 0.5 for both basins. Precipitation has a smaller control, and the precipitation-temperature interaction has negligible to minor influence (Figure 4c,d). The higher temperature control on Q min is also illustrated by stratified patterns along O-M_T axis, i.e., larger Q min with higher O-M_T, especially in the case of Liard-LC basin. The MLR model predictability for Q min as given by R 2 range between 0.39 and 0.46; thus, the models can only explain <50% of the variance of Q min . Nevertheless, the higher temperature control on low flow-which in this region is mainly contributed by baseflow from groundwater [52]-align with the expectation of higher groundwater recharge and subsurface flow and higher Q min , as temperature increases [20].
In contrast to Q min , precipitation has higher control (VI 2 ≥ 0.59) on Q max , especially in the case of Liard-LC. This is also illustrated by a stratified positive relationship between Q max and O-J_P, i.e., higher Q max for larger precipitation (Figure 4e,f). The temperature control on Q max is ≤ 0.26, indicating the smaller role of O-J_T change on Q max change. In this case also, the precipitation-temperature interaction plays a minor role on Q max . The MLR model R 2 range between 0.46 and 0.57, indicating a moderate predictability of Q max with O-J_P and O-J_T.
Overall, the evaluation of precipitation and temperature controls on Q mean , Q min and Q max provide further insights into the effect of these variables on streamflow response. The dominant control of precipitation on Q mean and Q max , and temperature on Q min from this observation-based study is consistent with the modelling study that include historical and projected future simulations [14]. This provides some confidence in using the MLR model for diagnosing the future direction of streamflow change with respect to the projected precipitation and temperature changes. As such, future warming can be generally expected to lead to higher Q min , while precipitation increase can be generally expected to lead to increased Q mean and Q max . However, the MLR models, with the R 2 predictability scores ranging between 0.39 and 0.74, have limited skills in explaining the variability of streamflow components. Furthermore, given the lack of physical representation, the MLR model is considered not suitable for extrapolation problems, such as projecting future changes. Nevertheless, the MLR-based VI analysis contributes to a process understanding with regard to dominant driving variables affecting streamflow responses.

Discussion on Future Changes in Extremes
The preceding investigation on the relationships of precipitation and temperature with Q mean , Q min and Q max , based on the historical climate, leads to a subsequent question about the potential future changes in these variables in the context of enhanced warming and increased precipitation across the pan-Arctic region [53]. A number of other factors are also associated with the higher temperature and precipitation are that directly or indirectly affect these streamflow variables. For instance, with warmer temperatures, there will be a reduction in the snowfall fraction and an increase in the rainfall fraction [11]. The changing precipitation phase could lead to an increased frequency and areal extent of rain-on-snow events [16,18,54]. Future projections of snow indicate a general decline in the snow cover extent [7,55] and highly variable maximum snow-water equivalent changes, with increases, decreases or no change depending on the region, and climate model used [7,[56][57][58]. Furthermore, in response not only to warming, but also to changes in snow cover, which exerts a control on the underlying soil, Arctic permafrost is projected to undergo substantial degradation [59].
Given that the projected changes in temperature, precipitation, snow and permafrost are generally a continuation of historical trends, future changes in streamflow variables can be expected to follow the historical changes. For instance, previous global-scale or pan-Arctic-wide studies indicate future increases in Q mean and Q min , and earlier Q max timings [60][61][62]. Similar changes in Q mean and Q min were also projected for the Liard River basin [14]. However, the potential future changes in Q max over the northern region, which is conditional on the evolution of precipitation, temperature, snowpack and permafrost changes and their interactions, remain less clear. Nevertheless, there are indications of changes to the nonstationary streamflow extremes (i.e., alterations in return periods and return levels). For instance, Hirabayashi et al. [63] projected future increases in the frequencies of the historical 100-year return period floods for the Yenisei, Lena, Mackenzie and Yukon Rivers, and a decreased frequency for the Ob River.
Summer rainfall-driven peak flow events are also related to rainfall/snowfall balance, which are historically very rare in large Arctic rivers, and typically occur only in the small coastal catchments [64][65][66]. For instance, an extreme summer flood event occurred in the Upper Kuparuk River catchment (drainage area 142 km 2 ) on the northern slope of Alaska in response to a 50-h, 80 mm rainfall event in July 1999, with the peak flow exceeding snowmelt peak discharge for 1993 to 2001 [64]. Additionally, a closer examination of the Yukon-E streamflow revealed occurrences of multiple summer peaks, with secondary peaks in late summers of 2010 and 2012 (not shown). Given that portions of the Yukon Basin are experiencing extensive melting of alpine glaciers and perennial snow fields [67,68], these summer peak flow events may have been influenced by glacier melt. In this respect, the projected glacier mass loss has the potential to modify the mean and extreme streamflow states of basins with glaciers across the pan-Arctic, including changes in the summer peak flows [69].

Summary and Conclusions
This study contributes towards understanding the historical changes in streamflow regimes and the role of climatic drivers on streamflow changes across the continental scale permafrost region of Canada. We analyzed historical trends in several streamflow components (i.e., Q mean , Q min , Q max and Q max timing) across 84 streamflow stations in the permafrost region of Canada. Overall, the trends are generally consistent with previous studies across the pan-Arctic (Appendix A Table A1), with increasing Q min for 43% of stations (p < 0.10) as the most prominent trend. Q min trends are also field significant for 38% of stations, and the majority of the stations with Q min increases are located in western Canada. Similar to previous studies (Appendix A Table A1), the number of stations with significant trends in Q mean , Q max and Q max timing are small, with increasing or decreasing trends of less than 15%, and even smaller fractions with field significant trends. Temperatures over the basins that drain to hydrometric stations show widespread increases, with significant trends in October to March and April to September temperatures for 80 and 71% of the basins, respectively; the majority of trends are field significant. Upward precipitation trends were obtained for some of the basins, especially for the months of October to March, with 27% of basins showing increases and 13% field significant.
The analysis of climatic controls (precipitation and temperature) provides key insights into their relationships with the streamflow changes. For instance, the Kendall's τ correlations of the basin-averaged runoff with precipitation and temperature indicate weak to moderate associations, including positive correlations of q mean with O-M_P and A-S_P, and q min with O-M_P and O-M_T. These associations reinforce the role of higher basin precipitation in producing higher runoffs. In contrast, the negative correlations of A-S_T with q mean and q max is indicative of the temperature influence on evaporation, snow accumulation and melt; therefore, the higher the temperatures, the lower the runoffs. The correlations of streamflow trends with precipitation and temperature trends are usually weak, although directions of trends are generally in line with the expectation, i.e., the positive correlation between q min and O-M_P trends, and the negative correlations between q max timing and A-S_T and O-M_T trends.
The MLR-based variable importance analysis captured the sensitivity of climatic drivers on streamflow components and provided plausible explanations on driver-response relationships. The results revealed the dominant control of precipitation on Q mean and Q max , with increasing wetness generally leading to higher discharge. On the other hand, the dominant temperature control on Q min reinforces the impact of climate warming on increasing the Q min trend, apparently by increasing the fraction of precipitation falling as rain and promoting subsurface flow with increasing temperature. Overall, the insights gained from this study, i.e., the effect of climatic drivers on streamflow changes, could be generalized for understanding streamflow changes across the permafrost region of the world. Furthermore, with a careful selection of climatic drivers, the methodologies used in this study could be applied to evaluate streamflow changes in the permafrost-free regions.
The historical streamflow trends and their relationships with climatic drivers lead to a subsequent question on future streamflow changes. In this respect, previous studies suggest future increases in Q mean , Q min and earlier Q max timings across the pan-Arctic [14,[60][61][62], while the direction of Q max -which depends on the interactions of precipitation, temperature and snow-remain uncertain for most regions. Furthermore, given that most of the impact studies over the northern region are large-scale studies, and not designed to represent the finer details of the streamflow extremes, there is a need to develop basin-scale hydrologic models by incorporating key processes (such as permafrost degradation) for an improved understanding of the ongoing and future hydrologic changes. Furthermore, there is a need to improve the ground-based streamflow observation network as well as leverage remotely sensed products to better evaluate and attribute the streamflow changes, especially in the areas with sparse coverage as identified in this study. Overall, the results of this study, including the enhanced understanding of ongoing streamflow changes and climatic controls affecting these changes, provide valuable insights into the hydrologic processes in the permafrost region of Canada and beyond.  The review of streamflow trends across the Pan-Arctic generally suggests consistent patterns. For instance, in the case of Qmean, most studies found an increasing trend or no trend (p < 0.05 or p < 0.10), with no study showing significantly decreasing trends. Furthermore, studies that considered aggregated discharge from several stations also found generally increasing discharges. For instance, Zhang et al. [4] found an increasing trend for the combined annual discharge from Lena, Yenisey and Ob Rivers from 1948 to 2008. Likewise, Durocher et al. [26] found a general increase in freshwater flow to the Arctic Ocean for the period 1975 to 2015, with the increase being more prominent in the Eurasian rivers than in the North American rivers. Ahmed et al. [28] found that the combined annual mean flow from the Mackenzie, Lena, Yenisey and Ob rivers to the Arctic Ocean has increased by 14% during in period 1980 to 2009. The increasing trends in Qmean are also consistent with the annual cold season and warm season precipitation increases across the Pan-Arctic [27].
From the few available studies on Qmin, the trends are mostly increasing, some of which are significant [21,70]. Additionally, winter (e.g., December to February or January to March) flows are also significantly increasing [28,34]. In the case of Qmax, the changes are mostly insignificant, except for the tributaries of the Yukon and Lena rivers, which show either increases or decreases [48,70]. However, besides hydro-climatic controls, other factors may contribute to these trends. For example, the increasing trend for the Lena River is also influenced by reservoir filling that reduced its outflows in the 1960s  Table 1 (Kolyma, Lena, Yenisey, Ob, Mackenzie, and Yukon). Red dots indicate the location of two hydrometric stations: (1) Yukon River at Eagle (Yukon-E) and (2) Liard River at Upper Crossing (Liard-UC); for which detailed analyses of climatic controls on streamflow extremes were conducted.
The review of streamflow trends across the Pan-Arctic generally suggests consistent patterns. For instance, in the case of Q mean , most studies found an increasing trend or no trend (p < 0.05 or p < 0.10), with no study showing significantly decreasing trends. Furthermore, studies that considered aggregated discharge from several stations also found generally increasing discharges. For instance, Zhang et al. [4] found an increasing trend for the combined annual discharge from Lena, Yenisey and Ob Rivers from 1948 to 2008. Likewise, Durocher et al. [26] found a general increase in freshwater flow to the Arctic Ocean for the period 1975 to 2015, with the increase being more prominent in the Eurasian rivers than in the North American rivers. Ahmed et al. [28] found that the combined annual mean flow from the Mackenzie, Lena, Yenisey and Ob rivers to the Arctic Ocean has increased by 14% during in period 1980 to 2009. The increasing trends in Q mean are also consistent with the annual cold season and warm season precipitation increases across the Pan-Arctic [27].
From the few available studies on Q min , the trends are mostly increasing, some of which are significant [21,70]. Additionally, winter (e.g., December to February or January to March) flows are also significantly increasing [28,34]. In the case of Q max , the changes are mostly insignificant, except for the tributaries of the Yukon and Lena rivers, which show either increases or decreases [48,70]. However, besides hydro-climatic controls, other factors may contribute to these trends. For example, the increasing trend for the Lena River is also influenced by reservoir filling that reduced its outflows in the 1960s [71,72]. From the very few available studies, it appears that Q max timings are shifting earlier [47,72] although the trends are generally not significant. Nevertheless, there is a general expectation of earlier Q max timing, along with the earlier onset of snowmelt and earlier centroid of annual flow in a warmer climate [73]. Yukon and other rivers in Alaska Bennett et al. [48] 1954/1964-2013 +3 (8) −3 (8) ++ (−−) indicate increasing (decreasing) trends at 5% significance level; + (−) indicate increasing (decreasing) trends at 10% significance level. NS05 and NS10 signify no significant trends at the 5% and 10% significance levels, respectively. Blank cells mean no analysis for that variable is performed in the cited study. The numbers in the trend columns are the number of stations with statistically significant trends (out of total number of stations).