4.1. Cluster Analysis and Trends in Cluster-Mean Time-Series
The data reduction clustering analysis identified five clusters, shown in
Figure 3A; for comparison, clusters of sites with similar temporal variations in streamflow identified by McCabe and Wolock [
7] are shown in
Figure 3B. Sites included in cluster 1 (four sites) were in west-central Florida, cluster 2 (six sites) were in inland central Florida, and cluster 3 (ten sites) included sites in northern Florida and southern Georgia. The ten-site limit imposed by McCabe and Wolock [
7] created their cluster 12, which roughly corresponds in location and in size with the combined clusters 1, 2, and 3.
Clusters 4 and 5 were the largest clusters identified, with 55 and 64 sites in each cluster, respectively. Sites in cluster 4 (which corresponds roughly with McCabe and Wolock’s [
7] cluster 7) were in western Georgia and extended as far west as Louisiana. Sites in cluster 5 (which corresponds roughly with McCabe and Wolock’s [
7] cluster 6) were in western Louisiana and extended to the far western boundary of the study area.
The mean time-series Z-scores for each cluster were within a range of −1.44 to 5.77. Cluster 1 had the largest range of Z-scores (4.64), and cluster 4 had the smallest range (1.086). Median Z-score values were similar for each cluster, which indicated sites in each cluster were similarly distributed around the cluster time-series of seasonal mean Z-score (
Figure 4). Clusters 1, 2, and 3 each had relatively high Z-scores from 1957 to 1960, 1998, and again from 2003 to 2005, indicating streamflow was greater than one standard deviation from the long-term average. These instances typically occur over multiple seasons: Spring, summer, and fall during the period from 1957–1960, winter and spring in 1998, and throughout entire years from 2003 to 2005. In the winter of 1998, clusters 1 and 2 had unusually large Z-scores, possibly due to record rainfall in central Florida and other areas in the Southeast [
50].
The patterns observed in clusters 1, 2, and 3 correspond to the seasonal distribution of precipitation (mostly occurring in the winter and spring months). Clusters 4 and 5 generally had more variable Z-scores; however, the values were typically above the long-term average from 1973 through 2005. A similar pattern occurred in cluster 3 over the same period. The boundaries of the clusters derived from the cluster analysis are considered to delineate areas of relatively homogeneous climate patterns, including in the temporal variation of climate, and therefore, to delineate areas within which trends in climatic drivers have produced similar trends in streamflow patterns. This result corresponds with McCabe and Wolock’s [
24] findings of a step increase in streamflow in the 1970s.
Results of the Mann–Kendall trend test of Z-scores of streamflow from each cluster of mean seasonal time-series data (1950–2015) indicated that clusters 1, 2, and 3 each had significant decreasing trends over the period of analysis, and cluster 5 had a significant increasing trend (
p < 0.05, α = 0.05). Cluster 4 did not have a significant trend (
Table 1). In addition, the magnitudes of the Sen’s slope values were small for all clusters, an indication of a very gradual trend (increasing or decreasing). The lack of a significant increasing or decreasing trend for cluster 4 may reflect an even balance between sites in the cluster with increasing and decreasing trends; its location between clusters 1, 2, and 3 to the east, which corresponds with cluster 12 from McCabe and Wolock [
7] (decreasing trends), and cluster 5 to the west, which corresponds with cluster 6 from McCabe and Wolock [
7] (increasing trend), reinforces even balance as a possible explanation.
4.2. Correlation between Seasonal Time-Series of Cluster-Mean Streamflow and Climate Indices
The five climate indices only explained a small fraction of the variability in the mean seasonal streamflow at sites within the clusters, which is a result that corroborated the findings in McCabe and Wolock [
24]. Streamflow increased at sites in the clusters as the value of the PDO index increased, and clusters 1 and 5 had the most significant correlation with the PDO (
Table 2). Correlations were positive and statistically significant (
p-values < 0.05, α = 0.05) between cluster time-series Z-scores and the PDO, except for clusters 4 (
Table 2). The correlation between all clusters and the PDO had similar values, which indicated that the effect of changes in the PDO on streamflow was similar for each cluster, albeit the correlations were weak.
Similar to the correlation with the PDO, the correlations between the seasonal cluster time-series Z-scores and the seasonal time-series ENSO index were statistically significant (p-value < 0.05, α = 0.05) in clusters 1–3 and 5. Similar results for ENSO and PDO were expected because the ENSO and PDO indices are both based on SSTs in the Pacific Ocean; however, the PDO is based on the long-term decadal mean of SSTs, whereas ENSO is based on the monthly mean. There has been a longstanding debate in the field of climatology about the relationship between ENSO and PDO and the mechanisms driving their respective modes, as well as their independence from one another.
In a study exploring the effect of ENSO on drought in the continental U.S., Rajagopalan et al. [
51] and Tuttle et al. [
52] described the PDO as having no significant moderating effect on ENSO over most of the continental U.S. Studies by Juanxiong et al. [
53] and Hamlet and Lettenmaier [
54] reached similar conclusions. The findings are in contrast to the findings of Gershunov et al. [
55] that determined the PDO was responsible for the modulation of ENSO-related winter precipitation in the U.S. Studies conducted by Allan et al. [
56] and Verdon et al. [
57] also concluded that the PDO and ENSO were related; however, the multi-decadal period of the PDO described several relatively short-term ENSO events. The results of this study suggest agreement with these findings [
55,
56,
57].
Sites located in the western and central part of the study area (clusters 4 and 5) and in northern Georgia (Cluster 3) were significantly correlated (negatively) with the AMO (
p-value < 0.05, α = 0.05). This result agreed with the results of Enfield et al. [
58] that described continental-scale patterns of negative correlations between rainfall and the AMO index. In contrast to sites in the eastern part of the study area (clusters 1 and 2) (
Table 2). The correlation between seasonal streamflow and the AMO was most significant at sites in Cluster 3. Over the study period, the mean of the AMO was slightly negative but near zero. Based on this result, streamflow in the clusters that were correlated with the AMO tended to vary inversely, resulting in increasing streamflow as the AMO index decreased. There was no significant correlation between the cluster time-series Z-scores and the NAO. The PNA index had a significant positive correlation with each cluster, except for cluster 4—which is likely a result of the wide-ranging, but subtle influence the PNA has on the ENSO, PDO, and AMO climate indices (
Table 2).
The climate indices explained approximately 8% or less of the variability in streamflow in the five clusters. The PDO explained from 1.1% to 4.26% of the variability in the cluster time-series Z-scores; the AMO between 0.3% and 7.63%, the NAO 0.26% or less, the ENSO between 1.23% and 5.3%, and the PNA between 0.07% and 3.97% (
Table 3). Because of the shared relationships between the climate indices and global SST, the cluster time-series Z-scores were tested against a field of mean seasonal global SSTs for the period of this study to identify potential teleconnections between ocean regions around the globe whereby SST and streamflow in the clusters were directly correlated.
Clusters 1 and 2 had similar patterns of spatial correlation with SST characterized by positive correlation along the west coast of North America and in the equatorial Pacific Ocean (
Figure 5;
Supplemental Material S1,
Figure S1A–E). Other areas of positive correlation included the British Isles, the mid-Atlantic, the Gulf of Mexico, the Caribbean Sea, and off the coast of Antarctica. Negative correlations between SST and streamflow for clusters 1 and 2 were generally found to occur in a band that wrapped around the globe between 30° S and 60° S. There was also a negative correlation with streamflow above 60° N in the Arctic Circle for clusters 1 and 2. The pattern of spatial correlation between streamflow in cluster 3 and global SST displayed the most correlation with oceanic regions around the globe. The pattern could be described as a more extreme version of the patterns observed in clusters 1 and 2.
Streamflow was positively correlated with SST in the northern and southern equatorial regions of the Pacific Ocean from the west coasts of Central America and South America in all clusters. A negative correlation between SST and streamflow was observed along the east coast of Asia, and in the waters off of the eastern and southern coasts of Australia, and in the South Pacific Gyre (cluster 1–4). There were also large regions of negative correlation that extended across the eastern seaboard of the U.S. and across the Atlantic Ocean to the British Isles. Cluster 4 was correlated with relatively fewer ocean regions around the globe. SST and cluster 4 streamflow were positively correlated in the Northern Pacific Ocean along the west coast of the U.S. and to some extent in the equatorial Pacific Ocean. Negative correlations were located in the central Pacific Ocean in the subtropics of the northern hemisphere and in the northern Atlantic near the British Isles. Cluster 5 was positively correlated with SST in more oceanic regions than what was observed with the other clusters. Streamflow from cluster 5 was positively correlated with SST along the equator, in the south Pacific Ocean, along the east coast of South America, and in portions of the Indian Ocean. There was a relatively small group of areas where cluster 5 streamflow and SST were negatively correlated. These areas included small regions of the North Atlantic and in the extratropic north-central Pacific Ocean.
4.3. Trends in Monthly Mean Streamflow
Maps depicting results from all monthly analyses can be found in
Supplementary Materials S1,
Figure S2A–L. For the period of analysis 1950–2015, March had the highest number of increasing trends (29), and May (4) had the lowest number of increasing trends (
Table 4). Overall, March had the highest number of statistically significant trends (44), and June had the lowest number (20). From 1950–2015, increasing trends outnumber decreasing trends in all months except February, April, May, and June. Visual analysis of January results shows a clear geographic difference between increasing and decreasing trends (
Figure 6). Decreasing trends are found in the eastern part of the study area, whereas all but two statistically significant trends in the western part of the study area are increasing. Analysis of the May 1950–2015 results indicate statistically significant decreasing trends cover the entire study area except for southern Texas (near Houston and San Antonio).
Of the four months where significant decreasing trends outnumber significant increasing trends, June trends are more balanced geographically with increasing and decreasing trends in both the eastern and western parts of the study area. In the western part of the study area (Texas), increasing trends are intermixed spatially with decreasing trends; however, in the east, as with the April and May results, decreasing trends are clustered around Tampa-St. Petersburg. From 1960 onward, statistically significant decreasing trends account for most of the monthly trends over all periods of record (
Table 4). During the multi-decadal periods of 1970–2015, 1980–2015, and 1990–2015, decreasing trends account for all statistically significant trends in the months of January, February, and May (
Table 4).
4.7. Spatial Variation in Streamflow Quantile Trends for Reference Sites (Reflecting Temporal Variation in Climate)
Trends at the 17 reference sites were significantly decreasing (
p-value < 0.05, α = 0.05) except for one site located in Louisiana (
Figure 10 and
Figure 11;
Supplementary Materials S1,
Figure S5A–D;
Supplementary Materials S2). The predominance of significant decreasing trends at all reference sites and for all trend periods during both the annual and seasonal time frames is striking. Decreasing trend slopes vary across the study area (slope percentage > 3% annually in Texas). The trend slope of reference sites in Mississippi and Louisiana (
Figure 10) indicates shallow decreasing trends (<2% change) for the longest trend period. The number of decreasing trends in the shorter trend periods indicate a significant decrease in streamflow beginning in 1980. The single exception is for site no. 81 (USGS station no. 07376500 Natalbany River at Baptist, Louisiana) for which increasing trends were detected for low streamflow percentiles for the longest trend periods, and in summer and fall only (
Figure 11).
For most reference sites in clusters 4 and 5, conditions were drier in 2015 than in 1970, but about the same as in 1950. The conspicuous lack of decreasing trends for the 66- and 56-year trend periods (periods starting in 1950 and 1960) despite the presence of decreasing trends for periods starting in 1970 and later is consistent with the theory that the one-time step increase in precipitation and streamflow that occurred rather abruptly near the year 1970 documented by McCabe and Wolock [
24] has acted to balance the strong and consistent decreasing trends in streamflow detected for the 46-year and shorter periods (starting in 1970 or later), so that no net change is detected when analyzing for the 66- and 56-year trend periods.
Although it seems contradictory that environmental factors that influence streamflow changes at reference sites—presumably climate or other atmospheric factors—would create these two opposing directions of change, it is possible that a change in one set of atmospheric factors in or around 1970 produced the step change, while a change in a different set of atmospheric factors has produced the consistent decreasing streamflow trends observed since 1970 (for example, see the discussion in Krakauer and Fung [
22]).
Spatial coherence of trends for the reference sites within a cluster improves when the distance from the coast is considered: For example, the reference sites (
Supplementary Materials S1,
Table S1) furthest from the coast in cluster 4 have fewer significant trends that are mostly confined to the 1970–2015 period, compared to reference sites closer to the coast. The proximity and possible spatial dependence (although on independent streamflow paths) of three of the five reference sites in the western part of cluster 5 are not ideal for evaluating spatial coherence for this cluster. As with cluster 4, distance from the coast may explain different patterns: Significant trends occur throughout the streamflow distribution and for three or more of the trend periods for the three sites in the western part of cluster 5, whereas trends are less common in trend periods with earlier starting years (before 1980) for the two reference sites near the coast.
The reference site sets are not ideal as streamflow patterns for paired reference sites on the same streamflow path, such as the Suwanee and Leaf Rivers (
Supplementary Materials S1,
Table S1), are likely to be autocorrelated unless intervening drainage renders them effectively independent. Because spatial correlation would invalidate conclusions about spatial coherence of reference site patterns, the sufficiency of intervening drainage area was analyzed to eliminate sites that could not be considered as independent observations; all reference sites were found to be sufficiently independent.
Several aspects of the trend results vary from east to west in the study area.
Steepness of trends: Clusters 1, 2, and 3 (
Supplementary Materials S2) have very few significant trends, but the slopes of the trends are steep (>2% change per year). In cluster 4, closer to the coast, the significant trends for longer trend periods tend to be shallow, whereas the slopes of significant trends for shorter, more recent trend periods (1980–2015 and 1990–2015) tend to be steeper. This is especially evident in the results of the seasonally stratified Q-K (likely due to the same factors that caused fewer significant trends for longer periods). In cluster 5, all slopes of significant trends are steep.
Number of significant decreasing trends in higher streamflows (
Supplementary Materials S2): Almost no trends in high streamflows were identified for sites in clusters 1, 2, and 3. Trends in high streamflow were present at some (5) sites in cluster 4, and in cluster 5 all reference sites show decreasing trends in high streamflows.
Occurrence of trends in summer and fall streamflows (
Supplementary Materials S2): Every reference site in cluster 5 had a significant decreasing trend in fall streamflow. Perhaps the dominance of precipitation (as high as 70% of the annual total) during warmer months and the resultant evapotranspiration combined with water use in this region makes streamflow more sensitive to changes in climate, hence steeper slopes, and more significant trends.
This analysis offers a more complete description than previous studies that have examined trend results only in mean values [
22] or in minimum/median/maximum values [
24]. The results illustrated in the stacked Q-K graphs were condensed to four metrics for each site and multidecadal period—the counts of quantiles with significant increasing and decreasing trends, and the mean value of trend slope for increasing and decreasing trends. For example, for site no. 79, Big Creek at Pollard, Louisiana (
Figure 9A), the count of quantiles for the period 1970–2015 is 0 increasing trends (mean slope = 0) and 97 decreasing trends (mean slope = −1.0). Although these simple metrics lack detailed information on the location of the trend in the streamflow distribution, they are more suitable for quantitative comparisons between sites and clusters. The spatial distribution of the four metrics is shown in
Figure 12A,B, for the periods 1950–2015 and 1970–2015, respectively. These two periods were selected to illustrate the influence on trend patterns of the step increase of streamflow near 1970 [
24]. Because of the step increase, trend results for the 1970–2015 period reflect a wetter starting condition than results for the 1950–2015 period, and this is the likely explanation for the disparity in the number of decreasing trends for the two periods. For almost all the reference sites in clusters 4 and 5, at least 50 of the 365 streamflow quantiles show a significant decreasing trend. Trend slopes were especially steep (more than a 3% decrease per year) for sites in northern Florida and Texas.
The number of quantiles at a site with decreasing trends increased, dramatically moving from east to west (
Figure 11). As evidenced by the trends in all periods, seen in
Figure 11. Over all periods of record, reference sites in Texas had more trends than reference sites in the east (Florida, Georgia, and Alabama). The average and standard deviation of counts for each cluster and the mean trend slopes for the period 1970–2015 are reported in
Table 7; these quantify the differences between clusters in the number of quantiles with trends during this period.
4.8. Spatial Variation in Streamflow Quantile Trends for Non-Reference Sites
Southern Texas and the Apalachicola River basin in western Georgia have the greatest density of sites at which more than half of the streamflow quantiles show a significant decreasing trend (
Figure 13). Two time periods in this analysis bracket the step increase in streamflow in 1970 [
24]—1950 to 2015 and 1970 to 2015. Reference sites for these two areas (clusters 4 and 5) also had a large number of significant decreasing trends for 1970–2015, and therefore, part of the pattern for this period could be climate -driven.
The darkest blue circles represent sites (
n = 23) for which more than half of the streamflow quantiles show a significant increasing trend (
Figure 13). These are scattered throughout the western and central part of the study area for the 1950–2015 period, and almost exclusively in Texas for the 1970–2015 period. Because almost no increasing trends were observed at reference sites for either of these periods, it is concluded that any trends at non-reference sites are likely due to anthropogenic changes, such as impoundments or diversions. The spatial trend from east to west is markedly different compared to reference sites (
Figure 12 and
Figure 13;
Table 7): Much larger average counts of decreasing trends for cluster 3 for all sites (148) compared to reference sites (17), and much smaller average counts of decreasing trends for cluster 5 for all sites (72) compared to reference sites (190).