Variation in post-smolt growth pattern of wild one sea-winter salmon (Salmo salar L.), and its linkage to surface warming in the eastern North Atlantic Ocean

Variation in circulus spacing on the scales of wild Atlantic salmon is indicative of changes in body length growth rate. We analyzed scale circulus spacing during the post-smolt growth period for adult one sea-winter salmon ( n = 1947) returning to Scotland over the period 1993 – 2011. The growth pattern of the scales was subjectively and visually categorized according to the occurrence and zonal sequence of three intercirculus spacing criteria ( “ Slow ” , “ Fast ” and “ Check ” zones). We applied hierarchical time-series cluster analysis to the empirical circulus spacing data, followed by post hoc analysis of significant changes in growth patterns within the 20 identified clusters. Temporal changes in growth pattern frequencies showed significant correlation with sea surface temperature anomalies during the early months of the post-smolt growth season and throughout the Norwegian Sea. Since the turn of the millennium, we observed (a) a marked decrease in the occurrence of continuous Fast growth; (b) increased frequencies of fish showing an extended period of initially Slow growth; and (c) the occurrence of obvious growth Checks or hiatuses. These changes in post-smolt growth pattern were manifest also in decreases in the mean body length attained by the ocean midwinter, as sea surface temperatures have risen.

also by increased numbers of circuli (Haraldstad et al., 2016). Thus, for example, from scale circulus measurements for Atlantic salmon smolts reared experimentally for 12 weeks, and at a range of temperatures and dietary regimes, Thomas et al. (2019) noted the rate of circulus deposition to increase with temperature. Circulus spacings did, however, show a complex interaction both with diet and temperature, and were widest at 10.5 C, narrower at 15 C and intermediate at 6 C.
The narrow spacings at 15 C arose as a result of the rate of circulus deposition being disproportionately high in relation to individual growth rate in body length. Notwithstanding the likely complexity of environmental temperature and prey availability influencing circulus formation, measures of both the numbers and spacing of marine circuli for wild juvenile Atlantic salmon (Salmo salar L.) in their first (= post-smolt) marine growth season have been interpreted as indicators of growth variation, with possible linkage to changes in sizerelated mortality at sea (e.g., Friedland et al., 2000Friedland et al., , 2006Friedland et al., , 2009Hogan & Friedland, 2010;McCarthy et al., 2008;Peyronnet et al., 2015).
Following river emigration of a smolt, the typical pattern of marine growth of wild Atlantic salmon during the post-smolt growth season is for intercirculus spacing generally to increase up to a socalled First Summer Maximum (FSM; Friedland et al., 1993;Peyronnet et al., 2015) followed by a progressive decline to a series of tightly spaced circuli comprising a visually obvious winter annulus. The number of annuli permits the determination of the sea age of the individual adult fish at migratory return to freshwater. Whilst the deposition of a single annulus per calendar year is typical of salmon and many teleost species, it is notable that the scales of some tropical or subtropical fish species show two annuli per year. For example, the cyprinid Labeo cylindricus in Mozambique shows two growth checks per yearin April and in Octoberwhich are coincident with the conclusion of annual spawning and the occurrence of the physiological winter (Weyl & Booth, 1999).
Within the context of a typical pattern of once-yearly deposition of annulior "checks" of tightly spaced circulion teleost scales, it is recognized also that "false checks" (sensu Ottaway & Simkiss, 1977) can form on the scale during the active growth season. For species reared in aquaculture, changes in husbandry regime (e.g., movement of fish between ponds or switches in dietary feed) can result in the formation of a series of tightly spaced circuli similar to an annulus (e.g., Ibañez et al., 2008). Bilton and Ricker (1965) first reported the occurrence of false checks on the scales of wild Pacific pink (Oncorhynchus gorbuscha) and chum (O. keta) salmon of known age.
With respect to Atlantic salmon, the incidence of false checks on the scales of wild adults returning to freshwater across a range of years has been explicitly reported only by MacLean et al. (2000). Dating back to 1963, Marine Scotland Science have compiled detailed monitoring data and scale sample archives for multiple river stocks of Atlantic salmon returning to Scotland. From these, MacLean et al. (2000) found that whilst the occurrence of false checks during the first (post-smolt) growth season was sporadic up until the 1990s, their incidence in the late 1990s showed a significant increase. Those observations, coupled with previous studies linking variation in post-smolt growth to long-term changes in year-class survival (e.g., Friedland et al., 2000Friedland et al., , 2009Peyronnet et al., 2015), highlight the potential importance to salmon stock management of a more detailed appraisal of marine growth variation (manifest as observable overall growth patterns or variation in zones of inter-circulus spacing) at the level of the individual fish. Here, we focus on variation in growth pattern based on circulus counts and spacings comprising the post-smolt growth season for time-series samples (1993-2011Marine Scotland Science) Tattam et al., 2003), and in roach (Rutilus rutilus; Britton, 2010). For Pacific and Atlantic salmonids, DFA has aided in the allocation of individuals to particular stocks and also the continent of origin for fish taken in mixed-stock fisheries (reviewed by Reddin & Friedland, 1999). In contrast to DFA, we here apply a hierarchical time-series clustering methodology to measures of variation in postsmolt marine circulus spacings up to the ocean winter annulus, with a view to objectively assessing the validity of subjectively and visually identified pattern categorizations for individuals. The fundamental difference between hierarchical time-series clustering and DFA lies in the latter relying upon previous knowledge of groups (i.e., "pattern") and group memberships, and on the overall approach to shape-based classification. The exploratory hierarchical clustering approach applied here considers the shape of the entire time-series sequence in assigning cluster membership, whereas the discriminant analysis classifier would account for scale characteristics expressed through different predictor variables. We show that hierarchical time-series clustering has the potential for interpretable pattern discovery in Atlantic salmon populations.

| Scale samples and circulus data
The present data pertain to return-migrant 1SW adult salmon captured in commercial nets in the estuarine reach of the River North Esk between 1993 and 2011. Marine Scotland Science personnel monitored catches throughout the netting season, which terminated on 31 August in each year. Monitoring of catches was at least weekly and all fish captured on a given day were measured. The present data are for stratified subsamples (n = 1947; 96-114 fish per year) spanning the full available date range of 1SW return migrants captured within each year (typically April/May to end August). Fish were measured (fork length, rounded down to 0.5 cm) and weighed (to 0.01 kg), and scale samples removed from the standard region (Hanson et al., 2019;Shearer, 1992). Scales were air-dried in paper packets and stored for further analysis. Scales were pressed on acetate strips and a single impression was photographed with a Leitz Wild M8 TODD ET AL. 7 FISH microscope; image analysis of circulus spacings was undertaken using AnalySIS software (Soft-Imaging Software GmbH; Münster, Germany).
Images were calibrated with on-screen accuracy ranging from 3.16 to 6.57 μm per pixel, according to scale radius. For each image the single ocean winter annuluscomprising the typical series of tightly spaced and slightly thickened circuliwas identified, and all marine growth circuli from river emigration of the juvenile smolt to the conclusion of the ocean winter annulus were enumerated and measured. The winter annulus amongst the present samples comprised a median of four circuli (IQR 3-5; range 2-11 circuli). An estimate of the body length of each fish at the midpoint of the winter annulus was derived by backcalculation of scale measures, and by applying the correction factor of Hanson et al. (2019) to the estimate of body length of the smolt at river emigration. Growth in body length between smolt emigration and capture of the measured adult fish was assumed to be isometric with the scale radius.
The primary objective of the present study was an analysis of variability and heterogeneity of the pattern of individual growth during the first (post-smolt) growth season at sea, as indicated by variation in the spacings of the pre-annulus circuli. Accordingly, a subjective assessment was made for each scale whereby the overall growth pattern of post-smolt intercirculus spacing was visually classified according to the zonal incidence of three criteria ("Fast", "Slow" and "Check" sequences). A sequence of widely spaced circuli was classified as "Fast" (F), more closely-spaced circuli as "Slow" (S), and circuli that were thickened and spaced similarly tightly to those of the subsequent winter annulus were allocated as a "Check" (C). Example scales showing the presently recognized SF, SFCF and F growth patterns are illustrated in Figure 1a-c, respectively, of Todd et al. (2014). Our recognition of a post-smolt growth Check is in accordance with the identification of "false checks" as reported by Ottaway and Simkiss (1977).
Here, when visually attributing a post-smolt (pre-winter) growth pattern (e.g., SFCF) for individual fish, the Check grouping was allocated only if there were three or more contiguous circuli comprising the feature. In some instances, single or double spacings of "Fast" or "Slow" circuli separating distinct zones were recognized. For example, for a scale showing a continuous series of widely spaced circuli, followed by a late Checkwhich itself was separated from the winter annulus by a single wide intercirculus spacethat fish was classified as FC. One especial challenge for some individual fish was the differentiation of the winter annulus circuli from a series of Check circuli that immediately preceded the start of the annulus (e.g., fish that were visually allocated a pattern of SFC). Here, we adopted the same protocol as Todd et al. showing the largest number of pre-annulus circuli. This standardization permitted the application of different clustering parameters, with the assurance from previous studies (Ratanamahatana & Keogh, 2004) of unaltered accuracy under reinterpolation, for the longest, shortest or mean length of clustered time-series. In order to achieve clusters based on similar shapes, a z-score normalization also was performed on each sequencefollowed, as part of the process, by a locally estimated scatterplot smoothing (LOESS) (alpha = 0.38)to highlight the underlying trend of each sequence without over-smoothing (Sardá-Espinosa, 2017).

| Hierarchical clustering
In applying time-series clustering, we explored several parameter settings for hierarchical clustering; all statistical analysis was undertaken using the computing environment R and the dtwclust package (R Core Team, 2018;Sardá-Espinosa, 2017. We undertook agglomerative hierarchical clustering, whereby every sequence starts as its own individual cluster and is sequentially grouped at each step of the algorithm according to the between-cluster and inter-group similarity (Hastie et al., 2009). Inter-group dissimilarity is known also as linkage and, as for other parameters, the appropriate linkage method is unknown a priori. As such, four linkage methods were explored: Simple, Average, Complete and Ward's.
For the analysis of the z-scored data, we utilized the two most frequently applied distance functions in other domains: dynamic time warping (DTW) and Euclidean distance (Silva et al., 2018). Euclidean distance often is used when measuring similarity in time-series clustering (Keogh & Kasetty, 2003). Clustering with Euclidean distance of z-normalized data has been shown to outperform clustering of the non-normalized counterpart, reinforcing the data preprocessing strategy adopted here (Tsumoto & Hirano, 2004). Although frequently used, and computationally efficient, Euclidean distance is not without its shortcomings. Unlike Euclidean distance, DTW is a distance measure capable of matching observations at different time points in order to achieve optimal nonlinear alignments, albeit with certain constraints (Silva et al., 2018). Although computationally intensive, DTW can provide nonlinear time-scaling invariance when comparing series that are locally out of phase ("warping") (Silva et al., 2013).
The dendrogram generated by hierarchical clustering will result in a different number of clusters (k) depending on the height of the cutpoint: the height of each cluster node is proportional to the value of the intergroup dissimilarity between its daughter nodes (Hastie et al., 2009). We have no previous knowledge of the cut level necessary to form an ideal number of clusters, but an interpretable k can be inferred from the relative heights of the dendrogram branches and their compactness and differentiation in shape. Following initial assessment of the overall cluster dendrogram that was generated, and given that the time-series data span 19 years, we chose to analyze in detail the variability and heterogeneity of the patterns of individual growth by partitioning into k = 20 clusters. The R code reproducing the data preprocessing steps of scale circulus spacing reinterpolation, z-scoring and LOESS smoothingtogether with the hierarchical clustering of the processed data, the cluster dendrograms and the cluster membership heatmapsare available at https://bitbucket.org/ arpmarquesm/scales-growth-patterns/src/master/. Also available in this repository is a sample of the data input to the R code for the three scales. The full data necessary to reproduce the results in this paper are available upon request.

| SST anomalies and correlation analyses
Sea surface temperature (SST) anomalies for the Norwegian Sea were derived from the NOAA OISSTv2 1x1 C gridded data set (www.cdc. noaa.gov). We took the monthly averaged data for each gridbox and show varying degrees of autocorrelation this had to be accounted for prior to the assessment of correlation significance. We did so by applying the method of Pyper and Peterman (1998) in calculating the required reductions of the degrees of freedom (from a maximum of 17) that were applied to the correlation coefficients. For the Fast data, the applied degrees of freedom ranged from 17 down to 9, whilst for the All Check data these ranged from 17 to 13.

| DISCUSSION
Despite the curtailment or closure of many high seas and coastal targeted salmon fisheries, the abundance of Atlantic salmon throughout the North Atlantic region has declined markedly over the past four decades (e.g., Chaput, 2012;Gregory et al., 2018;ICES, 2018). With specific reference to Scotland, a recent analysis of the ICES Working Group on North Atlantic Salmon (ICES, 2018) reported that the general decline in prefishery abundance of maturing salmon was forecast to continue. These declines likely reflect falling survivorship at sea F I G U R E 4 Time-series changes in ocean surface temperature and Salmo salar scale growth pattern. (a) Changes in monthly SST anomaly for the 250 and 500 km standard deviation spatially weighted kernels in the Norwegian Sea (April 1992 -March 2011). (b) Changes in frequency (proportion within years) of selected growth patterns. The three selected pattern groupings illustrate fish showing persistent Fast growth (F) throughout the post-smolt growth season, Slow growth followed by Fast growth (SF), and all patterns pooled that displayed one or more growth Checks. The growth pattern data for each capture year (b) are aligned with the SST anomaly in April of the previous year (a), coinciding with the commencement of annual smolt emigration (Olmos et al., 2019), but it is important to note that run-timing of returning 1SW adults in Scotland has shown progressive delays since the turn of the millennium (Todd et al., 2012) and that marine growth amongst European stocks also has been compromised in recent decades (Bacon et al., 2009;Bal et al., 2017;Jonsson et al., 2016;Todd et al., 2008).
The perception is that these changes in growth performance and phenology are not a direct physiological response of salmon to a warming ocean climate, but are indirectly driven by ocean warming and manifest as changes in the availability of epipelagic prey to salmon at sea (e.g., Beaugrand & Reid, 2012;Jonsson et al., 2016;Nicola et al., 2018;Piou & Prévost, 2013;Todd et al., 2008). Any such changes in prey availability to salmon may be pervasive throughout the duration of the marine migration, or perhaps they may present as marked seasonality and/or spatial heterogeneity. Hence, it can be argued that detailed appraisals of the variation in individual growth performance from analysis of scale intercirculus spacings, their temporal patterns and linkage to physical aspects of the exploited environment ought to prove additionally informative in furthering our understanding of factors underlying the recent decline in survivorship and growth of salmon at sea. Nonetheless, it has to be emphasized that data such as the presentobtained from time-series monitoring of home water commercial catches for an identifiable river stockare derivable only for those fish that successfully survived the marine migration to return as adults. We presently have no information for River North Esk fish that failed to survive.
The LOESS fits for all k = 20 dendrogram clusters ( Figure 1b) showed a variable, but qualitatively consistent, pattern of circulus spacing increasing to a maximum and then decreasing towards the commencement of the winter annulus. This is not dissimilar to the observations by Peyronnet et al. (2015) for 1SW adults returning to the Burrishoole catchment (western Ireland) between 1961 and 1999.
They noted intercirculus spacings generally to increase to a FSM and to decline thereafter toward the winter annulus. However, it is nota- Estimates of the rate of deposition of the earliest post-smolt circuli for wild salmon typically indicate an interval of approximately 6 days (Jensen et al., 2012;Todd et al., 2014). Accordingly, a smolt entering the sea in early June would probably deposit up to 10 fewer circuli over the post-smolt marine growth period compared to a smolt emigrating in early April. Should those two fish encounter identical marine growth conditions (and hence pattern of marine circulus spacing) throughout the post-smolt period, then the June emigrant will show an earlier FSM because it was "missing" the early marine circuli deposited by the April emigrant. It is notable that, for example, all nine clusters in sub-branches B and E (Figure 1b) are characterized by the peak occurring within the first half of the sequence. Conversely, for subbranch C (clusters 7-11) the peak was intermediate, whilst for clusters 2-5 of sub-branch A the peak occurred during the second half of the circulus sequence.
The contrasts between the present results and those of Peyronnet et al. (2015) probably are largely explicable by consideration of over-representation of our visually allocated growth patterns across the dendrogram (Figures 2, 3). For clusters 2-5 of sub-branch A we recorded a consistent over-representation of the SF pattern, manifest as a strongly negative skew of the cluster silhouettes. By contrast, sub-branch E (clusters 18-20) showed significant over- Monthly correlations between the SST anomalies throughout the post-smolt Salmo salar growth period and annual frequency of the Fast (F) and All Check growth patterns. The salmon data were lagged by -1 year to match the annual post-smolt growth seasons to the SST anomalies. Significant correlations (P < 0.05; following adjustment of d.f. to allow for autocorrelation) are shown by the filled circles The converse outcomeof a reduction in the frequency of the F growth patternalso is apparent. But it is important to emphasize here that these two groupings together account for a high proportion  (Figure 1a), of the 17 occurrences of significant over-representation of patterns including a Check sequence (Figure 2), nine were for those including an intermediate Check (e.g., FCS or SFCF) and eight were for terminal Check sequences (e.g., FC or SFC) immediately prior to the winter annulus. In first reporting the incidence of marked growth "checks" on scales from Scottish salmon, MacLean et al. (2000) noted most to occur within the second half of the circulus sequence comprising the post-smolt growth season. This was observed also for the present data (unpublished observation), leading to positive skew of the LOESS fits for given clusters. Consideration of the methodology of Todd et al. (2014) for estimating calendar dates of circulus deposition would indicate these Checks typically to pertain to the months of September-November prior to the formation of the winter annulus. A Check sequence of three to six circuli during the second half of the post-smolt growth period would indicate a period of markedly reduced growth persisting for perhaps up to 3-6 weeks . Such a protracted growth hiatus implies a major constraint on foraging of salmon at sea. In this respect, we believe that the classification and categorization of visibly obvious scale growth Checks and dating of scale circulus deposition might be especially informative in identifying periodsor distinct geographic areas within the ocean migratory trajectorythat currently are especially critical to salmon at sea.
Irrespective of the subjective allocation of growth patterns, our application of hierarchical clustering has permitted the resolving of clear structure within the available time-series of circulus spacing patterns. Given the widespread occurrence of marked growth hiatuses ("false checks"; Ottaway & Simkiss, 1977) during the active growth season of teleost fish species, we would argue that the present methods offer a broadly applicable tool for analyzing environmental influences on fish growth. Within the present salmon time-series, the most marked temporal shift in post-smolt growth performance in the marine environment was for the latter four or five adult return years, and much of the variation within and between the identified dendrogram clusters was indeed explicable by the visually allocated pattern.
The indications are, therefore, of a major shift in the typical postsmolt growth trajectory which occurred amongst fish returning to Scotland in the later years of the time-series. That timing matches similar contemporaneous declines in marine growth of adult 1SW salmon from southern Norwegian rivers . Our data show that the incidence of a late hiatus (= Check) has increased, whilst more fish also encountered poor growth (= initial Slow) throughout the first few weeks or months of the marine migration ( Figure 5). We therefore believe that this analytical approach can readily offer an effective synthesis of variation in the growth environment experienced both by species of Pacific and Atlantic salmon at sea, and perhaps provide informative linkage to size-related mortality during the marine migration (e.g., Friedland et al., 2000Friedland et al., , 2009Peyronnet et al., 2015).