The hydrometric (streamflow) stations selected for this study were all designated as “active” (i.e. were currently monitored), “natural” (i.e. their flows are not managed), and either continuous or seasonal and shown as having more than 30 years of data in ECDataExplorer (Environment Canada, 2010) at the time the data were downloaded. No attempts were made to use a common window of years – rather all analyses used the entire period of record for each station. In trend studies, time periods are selected that are a trade-off between record length and network density (Hannaford et al., 2013). Many trend studies use a common period of years with an arbitrary measure of completeness such as 20 years of data in a 60-year period (e.g. Vincent et al., 2015) and rely on continuous data throughout the year so that measures such as annual mean flow, or specific monthly flows, can be assessed for trend. This generally means that only continuously observed sites would be included. The alternative approach used here includes data from a large number of seasonal and continuously observed sites, which are compared using only the data available from April until the end of October.

The locations of the hydrometric stations, the main river basins, and major tributaries are shown in Fig. 1. Given the northerly (Mackenzie) or easterly direction of river flow in the region, the hydrometric stations generally sample basin hydrology that lies to the south or west of the points shown. The number of continuous and seasonal stations in each of the larger river basins is given in Table 1. Three additional stations were purposely included: these were at Changing Cold Regions Network (CCRN) Water, Ecosystem, Cryosphere, and Climate (WECC) observatories, Marmot Creek, Alberta, Smith Creek, Saskatchewan, and Scotty Creek, Northwest Territories (Table 1), and including these stations provides a link between the spatial patterns reported in this study and intensive process-based CCRN studies (DeBeer et al., 2016, 2021). Streamflow data from a total of 395 stations (gauged basins) were available; 233 (59 %) were operated on a seasonal basis. Water Survey of Canada station numbers are here referred to as stationID. Basin areas range from 9.1 km2 (Marmot Creek) to 270 000 km2 (Liard River at the mouth) and station elevations range from 22 m (Anderson River below Carnwath River) to 2095 m (Mistaya River near Saskatchewan Crossing). The data set contains values from nested basins which may cause some correlations; the analyses performed do not require sites to be statistically independent. Individual stations were analysed for all periods for which data were available, but clustering and statistical analysis involving multiple stations were restricted to the data in the annual common time window of the year from 21 April to 1 November when both seasonal and continuous data sets were available. Plots of missingness and annual station densities of the data set are provided in Fig. S1.

Satellite imagery and derived spectral indices are valuable for assessing effects of environmental changes and the hydrological responses of the gauged basins; these methods allow determination of changes in vegetation, water bodies and snowcover for large areas (e.g. Hall et al., 1995; Hansen et al., 2013; Pekel et al., 2016; Su, 2000). Time series of spectral remote sensing indices were constructed using optical imagery from the TM sensor (USGS and NOAA, 1984) on the Landsat 5 satellite for each gauged basin. The satellite had a 16 d return period between 1985 and 2010 and acquired imagery for any location on the Earth's surface with a spatial resolution of 30 m. The sensor was selected for its spectral capabilities, which allowed evaluation of surface changes, and its long operation which best suits the length of the hydrological record in the gauged basins. It was chosen to avoid combining data from different satellites or sensors to maintain consistency in spectral response over the study period. Later satellites use different spectral bands.

All analyses were performed with R (R Development Core Team, 2014) using packages kendall (McLeod, 2015), CSHShydRology (Anderson et al., 2018), and dtwclust (Sarda-Espinosa, 2017, 2018). A threshold of 0.05 was used in tests of significance, and accordingly, 5 % was also used as an indicator that the number of trends exceeds the number expected by chance alone.

As the intention was to include data from as many stations as possible, the entire period of record from each of the 395 gauged basins was used, and only the window where data were available from every station during the year from 21 April to 1 November was analysed. Table 2 provides the starting dates of the 5 d period corresponding to each 5 d period number from April through November. Because the station periods of record were used, rather than a common period of years, it was not appropriate to compare the magnitudes of trends among the stations. Instead, the analyses were restricted to determining the existence of significant trends in individual 5 d periods from 21 April to 1 November (periods 23 to 61, Table 2).

The main panel in Fig. 2 (bottom left) uses colour to show the magnitude of the flow for each day of each year for the Bow River at Banff, AB (stationID = 05BB001), an example of a long and complete continuous streamflow record from a national park in the Canadian Rockies. This streamflow record was described in detail in Whitfield and Pomeroy (2016, 2017). The upper panel of Fig. 2 shows the minimum, median, and maximum values for each 5 d period and blue (red) arrows indicate periods where there are significant increasing (decreasing) trends in streamflow over the period of record using Mann–Kendall tests. The directions of significant trends (i.e. positive or negative) were determined and were used subsequently for clustering of change types. The panel on the right shows the time series of annual minimum, median, and maximum discharges. If there was a significant trend (Mann–Kendall τ, p≤0.05), the series was coloured (red for decreasing, blue for increasing); black for no trend. The function for generating these plots was from CSHShydRology.

The stations in this study were operated for differing periods of time and with differing operating schedules. Figure 3 provides an example for the Bow River at Lake Louise, AB (05BA001), a station upstream of the Bow River at Banff, showing that this station had continuous operation between 1910 and 1920, was discontinued from 1921 until 1963, operated continuously between 1964 and 1986, and then had seasonal operation between 1987 and 2013, hence records only exist during the warm season (Fig. 3). The seasonal trends shown in the upper panel are based on all years in which data were present in the 5 d periods. In the right-hand panel, trends over time in annual minima, medians, and maxima are based only upon the years with complete data, and this shows gaps in the time series. Trends of these types should be based only using years with complete records and are not addressed further since a large proportion of the stations have seasonal operation. Figures S2–S13 show up to four example hydrometric stations for each streamflow regime cluster, as described below. Many of these plots show stations where the operation has alternated between being seasonal and continuous, similar to Fig. 3. Figures S2–S13 also demonstrate the variation in the years of record between stations. The complexity of the data set results from historical budgetary and management decisions in the Canadian hydrometric program. Assessing hydrological regimes and trends in this data set requires approaches that are different from “standard” methods.

Trends in each of the 5 d periods for the annual common time window were determined for the period of record of each time series, using Mann–Kendall tests as described above, following the approach of Déry et al. (2009b) for examining trend magnitude for a fixed endpoint in time. Interpreting the results for any fixed time period may not be representative of a longer timescale (Hannaford et al., 2013). As these were comparing periods separated by 360 d, i.e. a resampled time series, autocorrelation was not expected, and therefore pre-whitening was not applied. Tests with 100+ years of record show that nearly 90 % of cases did not show autocorrelation, and the balance was close to the level of significance (0.19). The individual station trend test results are indicated in the upper panels of Figs. 2, 3, and S2–S13, where significant increases and decreases are indicated by blue and red arrows respectively. The significant increasing, no trend, and decreasing trends were assigned scores of 1, 0, and -1 respectively. The individual annual trend scores for the annual common time window for the 395 stations were clustered using the method of k-means, which partitions observations into clusters having similar means and which is well suited to clustering of features such as patterns of significant differences (Likas et al., 2003; Steinley, 2006; Agarwal et al., 2016). The number of clusters chosen (six) was based upon the elbow method (Ketchen and Shook, 1996; Kodinariya and Makwana, 2013); using more than six clusters did not improve the modelling (not shown). These six clusters are hence referred to as “Trend Patterns”.

As the method used here is unconventional, we assessed how the patterns of trends would differ when using varying periods of record. Trends in each of 39 5 d periods were first determined for the entire period of record and were then compared to those of 21 periods, with lengths decreasing by 5 years between 1905 and 2005 (e.g. 1905–2015, 1910–2015, 1913–2015). Eleven stations with more than 3 years of data in the final interval (2005–2015) were excluded so that missing values were not introduced. The ∼ 15 000 comparisons (384 sites for 39 5 d periods) were used to determine the number of cases where the trends did not change. These comparisons were also tested using a classification approach and the adjusted Rand index (Morey and Agresti, 1985) and produced similar results (not provided here).

Given the large study region and the long study period, analyses of time series of Landsat 5 TM data are very resource intensive, both in terms of storage and computational power, and are beyond the scope of desktop computing. Google Earth Engine (GEE) allows for cloud-based planetary-scale analysis, while it serves as a database for petabytes of open-access satellite imagery such as the Landsat archive (Gorelick et al., 2017) and is particularly capable for this study.

Using GEE, Landsat 5 TM image composites (n= 579) were produced to cover the spatial extent of all the gauged basins for the period between 1985 and 2010 by mosaicking the available image scenes for each consecutive 16 d period. Theoretically this would allow for spatially complete mosaics given the 16 d revisit time of the satellite. To determine accurate spectral indices at the basin scale, pixels containing clouds or cloud shadows in each image scene were masked prior to mosaicking using the GEE-integrated Fmask algorithm (Zhu et al., 2015), which introduced intermittent data gaps to the set of mosaics. Additional data gaps were caused by occasional unavailability of satellite image scenes in the Landsat 5 TM catalogue, typically due to image quality or georeferencing issues (USGS and NOAA, 1984). To reduce the computation cost and data volume, the final mosaics were generated at an aggregated spatial resolution of 300 m. The total number of satellite image scenes used was 83 381.

For each basin, three time series of spectral index averages were derived from the 16 d mosaics of Landsat 5 TM data. Each normalized difference index (I) compares two wavelength ranges (W1 and W2) observed by the satellite detectors, using the form I = (W1 - W2)/(W1 + W2), each index ranging between -1 and 1. The three common indices used were the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Normalized Difference Snow Index (NDSI) (Lillesand et al., 2014) and were calculated by 1NDVI=RNIR-RRed/RNIR+RRed,2NDWI=RNIR-RSWIR/RNIR+RSWIR,and3NDSI=RGreen-RSWIR/RGreen+RSWIR, where R is the dimensionless top-of-atmosphere reflectance, Green is TM band 2 (green light 0.52–0.60 µm), Red is TM band 3 (red light 0.63–0.69 µm), NIR is TM band 4 (near infrared 0.76–0.9 µm), and SWIR is TM band 5 (shortwave infrared 1.55–1.75 µm). Because of the presence of the masked cloud pixels and data gaps, the basin means were often only calculated from a fraction of the complete pixel set of the basin; this fraction was determined for every observed time step.

For each 16 d period, the mean NDVI, NDWI, NDSI, and fractional coverage were determined for each of 375 gauged basins for which a shapefile of basin boundaries was available (20 basin boundaries were not available from Water Survey of Canada). A sample data set is shown in Fig. S14. Time series of annual maximum, mean, and minimum were determined for each of the normalized different indices and their fractional coverages from these data sets. Since there were only ∼22 images (at most) available for each year in each basin, the entire year was used rather than the annual common time window so that the annual minimum, maximum, and mean would be comparable across the study domain.

Forkel et al. (2013) demonstrated the annual variability of NDVI time series and the effects of using different analysis methodologies. Verbesselt et al. (2010, 2012) and de Jong et al. (2012) used breaks for additive season and trends (bfast) to detect change, particularly phenological change, in satellite imagery; bfast iteratively estimates the time and number of abrupt changes within time series derived from satellite images. While this methodology has considerable appeal and has been used widely and successfully to assess change in target areas, it was impractical to apply bfast here as it is difficult to summarize multiple changes in seasonality, trends, and breakpoints for three indices across 375 basins. A simpler approach, testing for simple trends in the mean, maximum, and minimum indices using Mann–Kendall (McLeod, 2015), avoids the rich complexity possible with bfast but still illustrates that NDVI, NDWI, and NDSI changes were accompanying streamflow regime changes. The minimum NDSI values excluded zero values.

The Streamflow Regime Types from the 12-cluster solution are shown in Fig. 6. Each of the 12 plots contains a line for each gauged basin in that type, and the heavy dashed line, where visible, is the centroid of all members; the colour of the lines is based upon stationID. The x axis is time, shown in increments of 5 d periods; these periods are renumbered starting with 1 for period 23. The y axis is the z-score (standard deviation units); the series for each site was converted to a z-score subtracting the mean and dividing by the standard deviation. The differences in the shapes of the hydrographs between regime types are evident and demonstrate how the shapes of the members and locations within a Streamflow Regime Type were similar. The outlying individual cases (e.g. Streamflow Regime Types 6, 7, and 8) are also evident.

Example plots for each of these 12 Streamflow Regime Types are shown in Figs. S2 to S13 to illustrate the similarity and differences between the hydrographs within the streamflow regime types. Streamflow Regime Type 1 basins were generally Rocky Mountain basins that have strong snowmelt and spring rainfall maxima signals (Fig. S2). Streamflow Regime Type 2 basins were reflective of Prairie streams with spring snowmelt and long periods with low or zero flow due to summer water deficits, variable contributing areas, and frozen conditions in winter (Fig. S3). Streamflow Regime Type 3 basins were in the Athabasca River basin dominated by humid upland boreal forest and lowland muskeg (Fig. S4). Streamflow Regime Type 4 (Fig. S5) has both strong snowmelt and late-summer streamflow. Streamflow Regime Type 5 basins were predominantly in the sub-humid Boreal Plains mixed forest (Fig. S6). Basins of Streamflow Regime Types 6–8 and 10 were unique (or nearly so) but were similar to adjacent types (Figs. S7, S8, S9, and S11). An interesting feature of these four types is that peak flows occur at different times in different years. Streamflow Regime Type 9 basins peak later in the summer and were generally smoother than for other basins (Fig. S10). Streamflow Regime Type 11 basins have an early snowmelt peak and high flows extending through into the fall (Fig. S12). Streamflow Regime Type 12 basins have an early snowmelt peak and persistent high flows during summer and extend into the fall (Fig. S13).

The spatial extents of the 12 Streamflow Regime Types were mapped over ecozones in Fig. 7; there is a clear spatial organization rather than a random pattern. This association is also evident in Table 3. Two large-scale features are evident: similar types tend to be from the same spatial areas, and some similar types follow along major rivers. Streamflow Regime Types 3 and 11 follow along rivers and Types 2 and 5 overlap. Streamflow Regime Types 1 (104 members) and 5 (148) occur in the greatest numbers of ecozones (8 and 6 respectively; Table 3).

Streamflow Regime Type 1 basins occur in the Cordillera (Montane n= 31 basins, Boreal n= 11, and Taiga n= 1) and also the Boreal Plains (n= 31), Taiga Plains (n= 5), Boreal Shield (n= 13) and Prairies (n= 11); the Type 1 hydrograph shows a sharp, brief melt period followed by a long, slow recession (Fig. 6). Streamflow Regime Type 5 basins were also common in the Prairie ecozone (n= 49), in the Boreal Plains (n= 81) as well as along the Mackenzie River to below Great Slave Lake; the hydrograph for this type shows an earlier and briefer peak than Type 1, with a rapid recession (Fig. 6). Streamflow Regime Type 4 basins (n= 10) were predominantly found in the Montane Cordillera in the west (n= 5) and in the Boreal Shield and Boreal Plains in the east (n= 2 each); the hydrograph shows prolonged high flows during the melt period and a short recession with relatively large flows (Fig. 6). Streamflow Regime Type 3 basins (n= 22) appear in the Boreal Plains (n= 17), Boreal Shield (n= 2) and Prairies (n= 3) and demonstrate the persistence of a mountain runoff signal along the Athabasca River, as this hydrograph contains the late-melt signal from glaciers and high-elevation snowpacks (Fig. 6). Streamflow Regime Type 2 basins (n= 85) were associated with the Prairie ecozone (n= 57) and Boreal Plains (n= 28), but this pattern also occurs in the Southern Arctic (n= 1) and Taiga Plain (n= 3); this pattern has the earliest snowmelt and most rapid recession, and the records often begin with snowmelt already in progress (Fig. 6). Streamflow Regime Type 11 basins (n= 12) were located near the mouth of the Mackenzie in the Taiga Plains and in the Hudson Plains along the Nelson River in Manitoba. Streamflow Regime Type 12 basins (n= 4) were in the Boreal Cordillera. Streamflow Regime Type 6–8 basins (1 each) and 9 (n= 2) were located at the edges of ecozones. While these descriptions are explicit, there was overlapping of types in space (particularly Streamflow Regime Types 2 and 5) and cases where individual basins of a type occur quite separately from each other (Types 9 and 12), as is evident in Fig. 7.

The standardized streamflows plotted in Fig. 6 make it difficult to compare the Streamflow Regime Types; plotting the z-score centroids of each (Fig. 8) makes the comparisons simpler. The non-Prairie Streamflow Regime Types have approximately linear recessions with two slopes (Types 1, 3, 4, and 12) or more (Type 11). The slopes of the recessions are listed in Table 4. Typically, the first recession phase was steeper than the second phase – where there was one. In Streamflow Regime Type 11 the third phase was steeper than the second but not as steep as the first. After a rapid recession, Streamflow Regime Type 2 becomes nearly horizontal, probably due to Prairie streams typically having no base flows due to lack of groundwater contributions. In Streamflow Regime Type 5 the recession has two linear phases and also terminates in a horizontal section, which was again likely to be caused by the absence of base flows in many Prairie basins. These recessions appear to have only five values, two that were steep (-0.22 and -0.16) and two that were much flatter (-0.06 and -0.04) in addition to the zero slope (no slope) of Type 5.

A caution is warranted here. The shapes of hydrographs are controlled by the climate, hydrological processes and landscape predominantly in the area of the basin where most of the runoff is generated, and while association with an ecozone was useful, the ecozone is a generalization and does not always capture hydrological functions in the source areas. In many rivers, the source of runoff lies in the high mountains, and these patterns are transmitted along the downstream river course such as in the Mackenzie, Liard, Athabasca, and Nelson rivers. Such stations are not independent. Differences between Streamflow Regime Type 1 and 4 basins likely reflect low late-summer flows from parts of the Rocky Mountains. In areas where there was overlap between different Streamflow Regime Types, both similarities and differences exist between the basins.

Trend Patterns are based upon the statistical trends in 5 d flows discussed above. Figure 9 illustrates the trend results for one station, 05DA007 Mistaya River near Saskatchewan Crossing, which drains a heavily glaciated mountain basin that is undergoing deglaciation. In the case of Fig. 9 there were three periods (39, 40, and 46) in July and August with significant decreases; the other periods have no trends.

Figure 10 plots all significant trends for the 395 basins, with significant increases and decreases shown in blue and red, no trend in gray, and no available data in white. These data were ordered by the six Trend Patterns determined using only the data for the 5 d periods from 23 to 61; Fig. S15 shows the data order by stationID. Periods 1–22 and 62–73 were not included in the clustering but were plotted as they are also of interest.

Trend Patterns in Fig. 10 are presented in the order that clusters were formed, showing the distinctly different Trend Patterns 1, 2, 3, 4, and 6, while Trend Pattern 5, the largest group with more than 250 basins (64 % of the total), does not have a consistent organized change despite there being individual periods with increasing or decreasing trends. Trend Pattern 1 shows positive streamflow trends in most of the annual common time window, suggesting a general increase in wetness throughout the spring, summer, and fall. Trend Pattern 2 has positive trends, suggesting increased wetness after period 30 in early summer. Trend Pattern 3 has predominantly significant negative trends, with many more in periods 40–61 than in periods 23–39, suggesting decreases in late-summer and fall streamflow. Trend Pattern 4 has significant increases centred about period 35, suggesting a shift increased snowmelt and rainfall-runoff peaks in June. Trend Pattern 6 shows significant increases in the first periods and last periods of the window but not during the summer; this group of stations all have winter data and show increasing streamflow throughout the late fall, winter, and early spring periods.

The trends presented in Fig. 10 are based on 39 5 d periods using the available data seta with records of at least 30 years. The stability of the trend results with decreasing-length observation periods is demonstrated in Fig. 11, where the trends are compared for increasingly shorter time periods with the results for the entire period. The fraction of the approximately 15 000 reduced-period results that are the same as the complete-period results is greater than 0.75, even when the record length is reduced to 10 years. The mean fraction of sites showing significant trends detected in each time period is about 0.20 and is at a maximum at 35 years and decreases at shorter time intervals (Fig. 11). For a period of 10 years, 5 % of the cases show significant trends, as would be expected by chance alone (based on a p value of ≤ 0.05). The impact of reduced length affects the Trend Patterns differently. Trend Patterns 1 to 3 show greater changes in the fraction of significant trends than Trend Patterns 4 to 6 (Fig. 11).

Associating a Trend Pattern with a Streamflow Regime Type is complicated as there were different numbers of basins in the Streamflow Regime Types and Trend Patterns (Table 5). Trend Pattern 5, which lacks any pattern, was very prominent in most of the Streamflow Regime Types having many members. Again, the caution regarding rivers sourced in the mountains and propagating the upstream signal downstream in nested basins also applies to their Trend Pattern as well.

The six Trend Patterns of the 395 hydrometric stations are mapped to ecozones in Fig. 12. Supplement Figs. S16–S21 provide more detail on the individual Trend Patterns and the station locations. Trend Pattern 1 (n= 19, S16) basins were located at the eastern and western margins of the Prairies with two north of 60∘, all showing increasing streamflows throughout the entire year. Trend Pattern 2 (n= 22, S17) basins, with early summer increases, were located across the Prairies and in the eastern portion of the Boreal Plain. Trend Pattern 3 (n= 32, S18) basins, showing decreased streamflow, were located predominantly in the western portions of the Boreal Plain and on the eastern edge of the Montane Cordillera. Trend Pattern 4 (n= 50, S19) basins, with larger early summer peaks, were located across the Prairies largely along the northern edge adjacent to the Boreal Plains and in a few northern locations in the Boreal Plain. Trend Pattern 5 (n= 254, S20) does not show an organized change pattern in the annual common time window and was distributed across all the ecozones. Trend Pattern 6 (n= 18, S21) basins, with higher cold and cool season flows, were located entirely in northern areas in the Taiga Plains, Taiga Shield, and Taiga and Boreal Cordillera. Overall, 28 % of the basins show one of the four increasing Trend Patterns, and 8 % have the single decreasing pattern.

At the ecosystem scale, 51 % of basins in the Prairies exhibit a definite Trend Pattern, with 45 % showing one of the increasing patterns (Trend Patterns 1, 2, 4, and 6) and 6 % the decreasing Trend Pattern 3 (Table 6). In the Taiga, increasing Trend Patterns predominate, with 46 % of stations in the Taiga Plains showing increasing patterns and none with the decreasing Trend Pattern, 29 % of the Taiga Shield having an increasing Trend Pattern and 14 % a decreasing Trend Pattern, and all three of the Taiga Cordillera basins having increasing Trend Patterns (Table 6). The Boreal Shield and Plains have increasing Trend Patterns in 16 % of stations and decreasing Trend Patterns in 13 %. The Boreal and Montane Cordillera have increasing Trend Patterns in 11 % of basins, and only the Montane Cordillera had the decreasing Trend Pattern (5 %) (Table 6). Trend Pattern 6 only occurs in the northern portion of the study area, which is dominated by thawing permafrost. None of the stations on the Hudson Plains showed any Trend Pattern. These results indicate that the streamflow regime change has a spatial basis, influenced by the location and ecozone, rather than by the Streamflow Regime Type.

The spatial patterns of trends in the mean values of the three normalized difference indices are presented in Fig. 13. The spatial patterns of the trends in the maximum, mean, and minimum of NDVI, NDWI, and NDSI are provided in Figs. S22–S24 and are also summarized in comparison with Streamflow Regime Type (Table 7), Trend Pattern (Table 8), and ecozone (Table 9). The tables show the fractions of stations grouped by trends that were significant at p ≤ 0.05. In the figures significant trends (p≤ 0.05) are shown as red (decreasing) or blue (increasing) triangles, trends whose significance was ≤ 0.10 are shown as red or blue dots, and those with no trend are plotted in black. There was a stronger association of the trends in the three indices with spatial location and with ecozones than with Streamflow Regime Type or Trend Pattern. Frequently, the trends in vegetation, water, and snow satellite indices occur in a spatial domain that follows the margin between two or more ecozones (Figs. 13 and S22–S24).

Trend Pattern 6 (Fig. S21), with prominent winter increases in streamflow across the permafrost-rich north (Taiga and Southern Arctic ecozones), was originally described by Whitfield and Cannon (2000). This region also has significant increases in mean NDVI (Fig. 13a) and decreases in mean NDWI and NDSI (Fig. 13b). At this scale, one interpretation of the observed trends would be that warming has altered the seasonal pattern and depth of frozen ground and has increased winter flows, groundwater connectivity, and also the greenness of these basins, which suggests increased evapotranspiration (Figs. 13a and S22) and has reduced both the amount of standing water (Figs. 13b and S23) and the snow-covered period (Figs. 13c and S24).

Three other increasing Trend Patterns (1, 2, and 4) show different temporal patterns of change. These three patterns were observed on the Prairies and southern Boreal Plains. These three Trend Patterns have considerable spatial overlap (Fig. 12).

Trend Pattern 1 (Fig. S16) was common in the Manitoba portion of the Prairie and Boreal Plains ecozone, with greater streamflows throughout the period when data were available. This region has no significant changes in mean NDVI (Fig. 13a), nor in NDSI (Fig. 13c), but there were increases in NDWI at the western edge of the region (Fig. 13b). One interpretation is that changes in vegetation and snow indices have been subtle. Streamflows have increased without altering the greenness (and hence evapotranspiration) of these basins (Figs. 13a and S22) or the snow-covered period (Figs. 13c and S24), but there has been increased wetness on the western side of this region (Figs. 13b and S23). The increase in storage is indicative of higher precipitation in this poorly drained post-glacial landscape.

Trend Pattern 2 (Fig. S17) was common in the Alberta and Saskatchewan portion of the Prairie and the Saskatchewan and Manitoba portion of the southern Boreal Plains ecozone. This region has no significant trends in mean NDVI (Fig. 13a) but does display significant positive trends in mean NDWI (Fig. 13b) and NDSI (Fig. 13c). An interpretation is that the increasing trend in water storage (Figs. 13b and S23) may be the result of increases in precipitation, and the increase in the snow-covered period may be due to increased snowfall, which through snowmelt runoff would also increase water storage (Figs. 13c and S24), but the greenness of these basins has not been altered (Figs. 13a and S22).

Trend Pattern 4 (Fig. S19) was common across the Prairie and southern Boreal Plains ecosystems. The region has no significant trends in mean NDVI (Fig. 13a) and only a few increasing trends in NDSI (Fig. 13c), which were restricted to Saskatchewan, but with significant increases in mean NDWI (Fig. 13b) except in Manitoba. Again here, the greenness of these basins has not changed (Figs. 13a and S22), despite trends in indices showing an increase in wetness (Figs. 13b and S23) which may be the result of increases in precipitation, and an increase in the snow-covered period which may be due to increased snowfall (Figs. 13c and S24).

Trend Pattern 3 (Fig. S18) was the only change pattern with decreasing streamflows and was common across the Prairie and southern Boreal Plains ecosystems. Decreasing trends in streamflow were more prevalent in the latter portion of the observed period. The region in which this pattern was observed is concentrated in the western part of the Boreal Plain and into the western portion of the Prairie ecosystem. This region has seen significant decreasing trends in mean NDVI (Fig. 13a), and the overlap of the areas with Trend Pattern 3 basins corresponds well to those basins demonstrating significantly decreasing mean NDVI and without trends in NDWI (Fig. 13b) or in mean NDSI (Fig. 13c). The greenness of these basins (Figs. 13a and S22) may have decreased because the basins are drier in summer since there are no decreases in wetness (Figs. 13b and S23) or snowcover (Figs. 13c and S24).

Streamflow regimes are a useful way of considering seasonal hydrology (Bower et al., 2004). Unfortunately, the large numbers of standardized streamflows plotted in Fig. 6 make it difficult to compare the Streamflow Regime Types; plotting the z-score centroids of each (Fig. 8) makes the comparisons clearer. Each of the Streamflow Regime Type centroids have a dominant peak and a specific shape. The peaks may be narrow (Types 1 and 11) or broad (Types 3, 4, 9 and 12). Streamflow Regime Types 2 and 5 associated with the Prairies were not always sampled on the rising limb, because (a) the peaks are due to snowmelt, (b) the Prairies melt early in the year, and (c) the snowmelt period often occurs before the beginning of the annual common time window.

Sample hydrographs of individual stations from each type are shown in Figs. S2–S13. These similarities among the stations are due to the predominance of snowmelt as the source of streamflow in this cold region, and the differences are related to the combinations of landscape processes that mediate the melt (hypsometry, surface storage, groundwater, and glaciers). The classification produced by the algorithm is spatially reasonable in that stations from similar landscapes, hydrology, and climates were clustered together (Table 3). Stations which are nested along a single basin (e.g. Athabasca River, Type 3) were clustered together as the signal derived from the mountains propagates downstream. Distinctions between clusters draining differing terrains, which overlap spatially, such as at the boundary between the Cordillera and Prairies, are to be expected as the landscape gradients or differences are large. Similar results were found in Ontario (Razavi and Coulibaly, 2013). Three clusters (Streamflow Regime Types 6, 7, and 8) each contained a single member, and one (Type 10) had only two members. These Streamflow Regime Types are very different from those containing large numbers of members, and any clustering of hydrographs in this way must balance common characteristics against uniqueness. The use of DTW to address these issues in hydrology has been suggested previously (Ehret and Zehe, 2011; Ouyang et al., 2010; Mansor et al., 2018). Overall, the use of dynamic time warping overcomes the timing differences due to latitude and elevation. While median flows were used here because of the large number of stations, individual years could be used to develop a consensus clustering if a smaller number of stations were to be used.

Like the Streamflow Regime Types, the Trend Patterns also show strong spatial organization; one pattern of decreasing trends, four types of increasing trends, and a large group without trends (Table 5). There was no clear link between the Streamflow Regime Type and the Trend Pattern (Table 5). Basins in mountainous areas generally lack consistent patterns of trends (Trend Pattern 5), but some exceptions do occur. Many studies in British Columbia have shown late-summer streamflow declines (Leith and Whitfield, 1998; Jost et al., 2012; Fleming and Dahlke, 2014). This may be the result of hydrologic resilience (Harder et al., 2015) in mountain basins east of the Continental Divide.

The Trend Patterns have spatial distributions that appear to be unrelated to individual ecozones but, as would be expected from past descriptions of the predominant climate processes on the Prairies, drying in the west and wetting in the east (Borchert, 1950; Rosenberg, 1987; Luckman, 1990; Whitfield et al., 2020), and extending beyond any one ecozone. For example, over the Prairie ecozone (Fig. 12), there are three Trend Patterns with differing patterns of increasing streamflow. Trend Pattern 1 basins are located at the eastern and western margins of the Prairies in the partly forested “parkland” ecozone fringe (Figs. 12 and S16). Trend Pattern 2 basins are located across the Prairies but are concentrated in the more humid east and in the eastern portion of the forested Boreal Plains (Figs. 12 and S17). Trend Pattern 4 basins are scattered across the Prairies and occur along the southern margins of the Boreal Plains (Figs. 12 and S19). The decreasing Trend Pattern 3 basins are prominent in the foothills, boreal forest, and grassland regions just east of the Rocky Mountains across the Boreal Plains and just west of the Prairies (Fig. S18).

Although annual trends in climate and streamflow have been reported widely (e.g. Zhang et al., 2001; Peterson et al., 2002; Déry et al., 2009a; St. Jacques and Sauchyn, 2009; Shook and Pomeroy, 2012; Vincent et al., 2015, 2018), seasonal studies are less common, although they allow determination of process shifts at finer scales of analysis (Whitfield and Cannon 2000; Whitfield et al., 2002; Bennett et al., 2015; Auerbach et al., 2016). The focus of the present study was on runoff timing changes using methods originally developed by Leith and Whitfield (1998) and subsequent improvements (Déry et al., 2009b). Many of the changes reported here have been observed by others. Increased winter streamflow in northern Canada was reported by Whitfield and Cannon (2000) and others subsequently (St. Jacques and Sauchyn, 2009; Bawden et al., 2015). Timing shifts have been reported, particularly in the onset of the spring freshet (Burn, 1994; Westmacott and Burn, 1997) and smaller summer recessions (Leith and Whitfield, 1998). In the headwater basins of the Columbia River basin in British Columbia, the timing of spring snowmelt runoff only changed slightly, and summer flows were unchanged (Hatcher and Jones, 2013). The spatial extent of changes in basins on the Boreal Plains and Prairies are novel findings as studies in the past have seldom incorporated data from hydrometric stations with seasonal records. Using a classification of annual runoff patterns in streams across the Prairies and adjacent areas, Whitfield et al. (2020) determined that basins in the eastern portion of the Prairies were becoming wetter and basins in the west becoming drier, as would be expected (Borchert, 1950) and consistent with the increased streamflows reported for Smith Creek (Dumanski et al., 2015). But, no changes in precipitation or runoff of most seasonal and continuous streamflow records were detected in the Prairies (Ehsanzadeh et al., 2016). Several studies of precipitation in summer across the Prairies reflect similar pattern changes (Akinremi et al., 1999; Asong et al., 2016; DeBeer et al., 2016). In the US Great Plains, intensified drying and increased numbers and durations of low-flow periods and greater flow events of shorter duration were identified (Chatterjee et al., 2018).

The spatial clustering of Trend Patterns does not coincide with Streamflow Regime Types, or exactly with ecozones. The contiguous change regions are broad in space and span ecozone boundaries. This is indicative of changes that are taking place at scales different from those that generate runoff (Streamflow Regime Type) but are related to broad-scale changes that would be expected with changes in weather and climate patterns across the southern portion of the study area and with climate warming in the northern portion. These patterns were partially described by Whitfield and Cannon (2000), and they have been explored and reported on by others since (Burn and Hag Elnur, 2002; Woo and Thorne, 2003; Fleming, 2007; Janowicz, 2008; Bawden et al., 2015; Tan et al., 2018).

The primary interest in trends of normalized difference indices (NDVI, NDWI, and NDSI) was in determining whether the changes were driving or following observed streamflow Trend Patterns. Changes in land use and hydrology at the basin scale can drive streamflow regime changes (Fohrer et al., 2001; Woo et al., 2008). Satellite imagery is increasingly being used for basin-scale analysis, but these studies generally focus on a small number of cases in a small region (Coppin et al., 2004; Bevington et al., 2018; Soulard et al., 2016; Militino et al., 2018; Jorgenson et al., 2018; Lee et al., 2018).

The results presented here demonstrate several broad spatial patterns of change that warrant closer examination in the future. Four specific regions have results that demonstrate significant changes, or lack thereof, in hydrology and satellite indices. These four regions correspond to areas with changes in water storage reported in Rodell et al. (2018). These four GRACE trend regions are consistent with the results presented here. Rodell et al. (2018) suggested the regions result from the following mechanisms: (1) precipitation increases in northern Canada; (2) a progression from a dry to a wet period in the eastern Prairies/Great Plains; (3) a region of surface water drying in the eastern Boreal; and (4) a region of no change along the Rocky Mountains. Based on the results presented here, mechanisms different from those suggested by Rodell et al. (2018) should be considered.

In the Arctic, the patterns, magnitudes, and mechanisms of hydrological and ecological change are often unpredictable or difficult to separate from other drivers (Hinzman et al., 2005). In many disciplines, understanding the complexity of the Arctic is challenged by a lack of scientific knowledge, observational and experimental time series, and the technical and logistic constraints of Arctic research. In the Boreal, Ireson et al. (2015) indicate that the current monitoring of climate, hydrology, and ecology are insufficient for understanding or predicting the potential responses to either human activities or a changing climate. Even with inclusion of more than 200 stations that are observed seasonally, there are regions in the study domain, particularly north of 60∘, that have too few observations, and consequently the basins that are observed are often larger compared to those in southern regions.

It is important to note that climate signals, particularly for the Pacific Decadal Oscillation (PDO) and Arctic Oscillation (AO), were not considered here. Oscillation in the climate system can be manifest as tests on short time periods (Chen and Grasby, 2009; Hannaford et al., 2013). Bennett et al. (2015) showed some connections of streamflow in Boreal Alaska to PDO but not to AO. The interannual variability of the Liard River is correlated with the PDO (Rood et al., 2017). Within mountainous areas, such as the Cordillera, the hydrological response to the regional climate variability signal is likely to be modified by local factors including location, topography, and land characteristics (Thorne and Woo, 2011). Future work should consider adapting the methodology presented here to examining the relationships of streamflow to these climate signals (and others), by resampling the available station records including complete and partial-year streamflow records in relation to the climate indices.

Are landscape changes related to initial state, and are they a response to or a driver of hydrological change? Broad-scale studies examining streamflow trends and timing changes should employ multiple methods across different scales and consider regime-dependent shifts to better identify and understand changes (Bennett et al., 2015). Here, it was shown that there are several coincident and overlapping regions of change in streamflow and in satellite indices. Further studies, at a finer spatial scale, will be required to determine process drivers and responses.

The main premise of this study was to let the available data tell the story of hydrological structure and change in this large study domain. Including data from seasonal stations that only report a part of the year more than doubled the number of stations available for analysis, and most of these stations are located in the Prairie ecozone, which has often been omitted in studies because of the lack of continuous data. The structure of the available data is complicated, the data set contains streamflow records from entire years and/or only the warm season and for a variety of periods of years, and the sites are not evenly distributed across ecozones. Determining the magnitudes of annual trends was not appropriate with the approach used because of the inconsistency of years of data between sites. But, including the data from both seasonal and continuous stations and an annual common time window rather than only entire years, the results provide an interesting spatial story of trend direction. While there are timing shifts in many Cordillera basins, the changes are inconsistent between sites, suggesting a resilience to change in complex terrain as reported by Harder et al. (2015). In the north, winter streamflows are increasing, and these are reflected in changes detected in satellite indices. The western Boreal Plains basins exhibit decreasing trends in streamflow and are less green (NDVI). There was a complex pattern of changing streamflows across the Prairies, with some drying in the west and areas with different patterns of increased streamflow particularly in the northeastern Prairies.

The motivation for this study came from the NSERC Changing Cold Regions Network study of 2013–2018 (DeBeer et al., 2015, 2021) of Western Canada's rapidly changing cold interior. That study sought to integrate existing and new sources of data with improved predictive and observational tools to understand, diagnose, and predict interactions amongst the cryospheric, ecological, hydrological, and climatic components of this study area. The results presented here are already informing several science research agendas in Canada and internationally. The results also contribute to the Global Water Futures program of 2016–2023 (http://www.globalwaterfutures.ca, last access: 5 May 2021), which has an overarching goal of delivering risk management solutions to manage water futures in Canada in a time of unprecedented change.

The first step to managing future water changes is to understand those of the recent past and that are currently underway, and this study makes a strong contribution to that process. In particular, this study represents a step forward in addressing the complexity of hydrological change; there are many studies of individual basins which, when results are considered individually, tend to be more anecdotal than systematic. It is indeed simpler and easier when there are only a few cases to consider with common variables and record length, as in many studies, but with large numbers of basins the tools for dealing with significant changes are more limited. Sensitivity studies that assess the limits of partial-year analysis of hydrological structure and change are required and seem a logical next step.

Increased streamflows, particularly in winter, are taking place in the northern portion of the Mackenzie Basin. In these basins, and in the general region, mean NDVI has not changed, but mean NDWI and NDSI have decreased. Degrading permafrost resulting in increased winter streamflow is an important change, which has been observed to take place in this region and which may be driving the observed changes in NDWI. Decreasing snowcover, evidenced by decreased NDSI, may be reflected in the shift in the partition of rainfall and snowfall due to warmer spring transition periods.

Decreased streamflows occur in these basins. In this area in general, mean NDVI has significantly decreased, but along the western margin mean NDWI and NDSI have increased. Decreased NDVI was consistent with decreased streamflow; as basins become drier, vegetation will be impacted. Increases in mean NDWI and NDSI occur only at the western margin of this region, suggesting a zone with complex gradients and changes.

Increased streamflows in summer and fall are occurring across the Prairies, particularly in the east, and along the northern margin and the southern margin of the Boreal Plains. Three types of hydrological changes are evident with differing spatial locations, but the overlap between types suggests that there are multiple processes involved. In these basins, and in this area in general, mean NDVI has significantly increased, mean NDWI has increased, and mean NDSI has not changed. This is consistent with the higher rainfall ratio observed since the 1950s in the region, recent higher summer precipitation and both increased water storage and drainage reported in the northern and eastern Canadian Prairies. Basins in the eastern Prairie show increased streamflows through the entire period of the year that data are available. Some basins, which are centred in Saskatchewan, show increases only during summer and fall which are indicative of changes in the streamflow regime from nival towards pluvial. Some basins show only a narrow time of the year with increased streamflows, which was the largest of these three change patterns, and are found in the western Prairie and along the margin between the Prairie and Boreal Plains. In these basins, mean NDVI has not changed, but NDWI has increased in basins in Saskatchewan and Alberta, and mean NDSI has increased in Saskatchewan. These differences may reflect the differences in the three change patterns. The existence of the east-to-west gradient of these changes was predicted by previous climatological studies (Borchert, 1950; Rosenberg, 1987; Luckman, 1990).

The mountain basins in this study appear to be resilient to change. These basins demonstrate several hydrograph types but generally lack structure in Trend Patterns. Individually, these basins do show periods with increases or decreases in streamflow consistent with freshet timing changes, as has been reported elsewhere, but there was sufficient inconsistency among the basins to define a specific pattern. Basins in these ecozones do have trends in NDVI, NDWI, and NDSI, but the results are mixed. There are decreasing trends in mean NDVI in the Montane Cordillera and mean NDWI and mean NDSI in the Boreal Cordillera and Taiga Cordillera and increasing trends in mean NDVI in the Boreal Cordillera and Taiga Cordillera, and there are no increasing trends in these three ecozones for either NDWI or NDSI; decreasing trends are more prevalent in the north than in the south.

The approach using an annual common time window presented here combines many more stations and years of data and increases the number of stations available for analysis quite dramatically from previous fixed time window methods. Analysis of only the annual common time window demonstrates groups with common spatial patterns in streamflow regimes and groups with common streamflow trends. A common annual time window in the warm season provides sufficient information to adequately resolve regional streamflow patterns and seasonal streamflow trends.