Social survival: Humpback whales (Megaptera novaeangliae) use social structure to partition ecological niches within proposed critical habitat

Animal culture and social bonds are relevant to wildlife conservation because they influence patterns of geography, behavior, and strategies of survival. Numerous examples of socially-driven habitat partitioning and ecological-niche specialization can be found among vertebrates, including toothed whales. But such social-ecological dynamics, described here as ‘social niche partitioning’, are not known among baleen whales, whose societies—particularly on foraging grounds—are largely perceived as unstructured and incidental to matters of habitat use and conservation. However, through 16 years of behavioral observations and photo-identifications of humpback whales (Megaptera novaeangliae) feeding within a fjord system in the Canadian Pacific (primarily within Gitga’at First Nation waters), we have documented long-term pair bonds (up to 12 years) as well as a complex societal structure, which corresponds closely to persistent patterns in feeding strategy, long-term site fidelity (extended occupancy and annual rate of return up to 75%), specific geographic preferences within the fjord system, and other forms of habitat use. Randomization tests of network congruency and clustering algorithms were used to test for overlap in patterns of social structure and habitat use, which confirmed the occurrence of social niche partitioning on the feeding grounds of this baleen whale species. In addition, we document the extensive practice of group bubble net feeding in Pacific Canada. This coordinated feeding behavior was found to strongly mediate the social structure and habitat use within this humpback whale society. Additionally, during our 2004–2019 study, we observed a shift in social network structure in 2010–2012, which corresponded with environmental and demographic shifts including a sudden decline in the population’s calving rate. Our findings indicate that the social lives of humpback whales, and perhaps baleen whales generally, are more complex than previously supposed and should be a primary consideration in the assessment of potential impacts to important habitat.

geography, behavior, and strategies of survival. Numerous examples of socially-driven habitat 25 partitioning and ecological-niche specialization can be found among vertebrates, including toothed 26 whales. But such social-ecological dynamics, described here as 'social niche partitioning', are not known 27 among baleen whales, whose societies --particularly on foraging grounds --are largely perceived as 28 unstructured and incidental to matters of habitat use and conservation. However, through 16 years of 29 behavioral and photo-identification observations of humpback whales (Megaptera novaeangliae) feeding 30 within a fjord system in British Columbia, Canada, we have documented long-term pair bonds (lasting up 31 to 12 years) as well as a complex societal structure, which corresponds closely to persistent patterns in 32 feeding strategy, long-term site fidelity (extended seasonal occupancy and annual rate of return up to 33 75%), specific geographic preferences within the fjord system, and other forms of habitat use. 34 Randomization tests of network congruency and clustering algorithms were used to test for overlap in 35 patterns of social structure and habitat use, which confirmed the occurrence of social niche partitioning on 36 the feeding grounds of this baleen whale. In addition, we document the extensive practice of group bubble 37 net feeding in Pacific Canada. This coordinated feeding behavior was found to strongly mediate the social Patterns of habitat use and the factors that determine it are of fundamental interest in animal ecology [1]. 48 As species occupy and use habitats according to various needs, they are establishing a certain ecological 49 niche within a common domain of limited resources [2,3]. To persist in shared space, sympatric species 50 may differentiate modes of habitat use over time to secure resources [4]. Within a single population, 51 individuals may even vary their movements, prey preferences, and behavior in order to reduce niche 52 overlap and minimize intraspecific competition [5]. Such habitat partitioning, on both inter-species and 53 intra-species levels, facilitates the coexistence of functional groups, increases local biodiversity and 54 ecosystem complexity, and uses available resources more efficiently [4]. 55 56 In any given species, spatiotemporal distribution and ecological niches are the result of negotiations 57 between environmental dynamics and ecological interactions, as well as intrinsic biological factors such 58 as ancestral body plan, reproductive status, and feeding strategy [6][7][8][9]. Social relationships are also an 59 important determinant of habitat use in certain species, particularly within the context of innately social 60 behaviors such as mating and parental care. The geography and seasonality of habitat use is often closely 61 related to mating systems and reproductive calendars, e.g., sea turtles [10], marine iguanids [11], many 62 pinnipeds [12], and most seabirds [13]. Even within a single species, mating systems can change in 63 response to resource availability and habitat type (e.g., Equus africanus in North America, [14]). 64 65 Among the cetaceans, the influence of social-reproductive behavior upon habitat use has been observed in 66 both odontocetes (toothed whales, porpoises, and dolphins) and mysticetes (baleen whales). Examples 67 include nursery groups of dusky dolphins (Lagenorhynchus obscurus) who use shallower waters than 68 other social groups in New Zealand [15], sperm whale (Physeter macrocephalus) populations whose 69 seasonal geography is partitioned by sex and age [16], baleen whales who practice large-scale movements 70 between winter breeding areas and summer feeding areas [17], and specifically humpback whales Outside of a breeding context, such as foraging, the social component of habitat use may not be 75 recognized unless the species travels within stable family groups (e.g., sperm whales and killer whales, 76 Orcinus orca) or large pods (e.g., tropical dolphins) [16,[30][31][32]. But even for relatively solitary species, 77 foraging and other non-reproductive behaviors still occur within a social context [33]. Individuals with 78 similar preferences tend to share common spaces and interact regularly with one another, thus adding a 79 social dimension to their patterns of resource use. These interactions can be merely coincidental or 80 actively sought out and coordinated. An example of the latter is when groups of humpback whales gather 81 together and engage in bubble net feeding to capture schooling fish [34,35]. As a complex feeding 82 strategy that requires cooperation and learned behaviors [35], 'bubble netting' is a dramatic example of a 83 habitat use strategy with a strong social component. Over time, as socially intensive behaviors such as 84 bubble netting recur, the bonds that form among individuals can serve to reinforce the ecological 85 similarity of an in-group and exacerbate its differences from out-groups [36]. Again, this process of 86 differentiation could happen coincidentally, via simple attraction [37], or actively, via social selectivity 87 and shared learning [38]. 88 89 It is in this way that the partitioning of habitat and ecological niche within a population can come to fall 90 along social boundaries, and that social and ecological roles can become mutually reinforcing. We shall 91 refer to this mutually reinforcing overlap of social-, habitat-, and ecological-partitioning as social niche 92 partitioning. This process has been observed among various vertebrate groups, including terrestrial 4 153 154 Second, with the same field effort, we sought to characterize the humpback whale society within this fjord 155 system by answering the following questions: how prevalent are strong social relationships in the local 156 population, and how stable are these relationships within seasons and across years? Furthermore, is there 157 internal structure to the social community, and is this structure stable across years? 158 159 Third, we used these data on whale habitat use and sociality to determine the extent to which these 160 aspects of humpback whale life mediate one another within this habitat. To do so, we asked whether 161 aspects of the sociality and habitat use of individuals were strongly related, and if so, in what specific 162 ways. Finally, to test for social niche partitioning, we asked whether population-level patterns in habitat 163 use correspond to the internal structure and interannual stability of the social network. Throughout these 164 analyses, we paid particular attention to group bubble net feeding (an innately social behavior), its 165 relationship to other aspects of habitat use, and its influence upon the social dynamics of the local 166 population. December) (Fig 1). When humpback whales were detected, groups were approached with caution, all 180 individuals were counted, location and behavior noted, and identification photographs of the underside of 181 their tail flukes were collected with standard DSLR cameras and telephoto lenses, following established 182 protocols for this species (e.g., [82] During close observation of whale groups, we identified six primary behaviors: bubble net feeding (BNF) 218 was identified by rings of bubbles and feeding at or just below the surface. Milling-feeding was inferred 219 according to several factors: travel pattern was circuitous or repetitively back-and-forth within the same 1 220 km 2 ; dives were long (more than 5 minutes); and surface sequences comprised many breaths during which 221 the animal was uncommonly still at the surface, suggesting recovery from feeding activity at depth. 222 Traveling was indicated by swimming in one direction for a period of time greater than 30 minutes (or 10 223 minutes when observed from shore) in which fluking was rare. Resting was indicated by directed travel at 224 low speed, low breathing rate, lingering at or just below the surface, and rare fluking. Sleeping whales 225 were motionless floating at the surface, breathing only one to two times per minute. Robust behaviors 226 included breaching, pectoral or tail slaps, and other energetic surface displays. We used 'social' as a 227 designation to capture behaviors during group interactions that were not clearly tied to the other behaviors 228 and were clearly directed towards others in the group who were in close proximity. These included Site fidelity 237 We characterized population site fidelity based upon the tendency of humpback whales to occupy the 238 study area or to return to it over some period of time [91]. To do so, we used population-level and 239 individual-based metrics of within-season and interannual observations (see Supplementary Appendix 1 240 for complete details of the analyses in this section).

242
Population-level residency patterns within a year were examined using Lagged Identification Rates (LIR; 243 [92,93]), which depict the probability that an individual identified on any given day will be re-identified 244 = 1, 2, 3, …, max days hence. We set max to 230 days within the same year as = 0 (chosen because the 245 maximum duration of any field season was 223 days), used only lags with 10 or more paired 246 identifications to build the LIR curve, and obtained confidence intervals using 100  We applied a Kruskall-Wallis rank sum test to ask whether site fidelity metrics for known bubble net 264 feeders differed from the remainder of the identified population (i.e. those we never observed bubble net 265 feeding). We then used permutation tests to determine whether the LIR of bubble net feeders differed 266 with statistical significance from that of the remainder of the identified population. The prevalence and stability of social relationships were characterized using weighted association indices 285 and Lagged Association Rates (LARs). We selected the Simple Ratio association Index (SRI, [100]) 286 because 1) we lacked calibration data and 2) the biases inherent to the SRI are more predictable than to an 287 alternative such as the Half-Weight Index [101,102].

289
Stability of associations 290 We calculated LARs using SOCPROG 2.9 and the R package "asnipe" [103] to describe the temporal 291 stability of relationships over time ([92]; Supplementary Appendix 3). For these analyses, we subset our 292 catalog to those individuals seen 10 or more times and used the same sampling periods and maximum 293 time lags as in the LIR analysis above. We tested whether dyadic stability differs between whales that 294 practice bubble net feeding and those that we have not observed doing so. This test involved calculating 295 several LAR curves: one for all groups encountered, a second for groups that contained at least one 296 known bubble net feeder (these groups could also contain whales not known to bubble net), and a third 297 for groups that contained whales not known to bubble net feed. As with the LIR analysis above, we used 298 permutation tests to evaluate the significance of LAR observations and conventional decay model fitting 299 in SOCPROG 2.9 to identify the feasible social processes underlying the temporal stability of 300 associations.

302
Social differentiation 303 We used maximum likelihood approximation in SOCPROG 2.9 to calculate social differentiation (S): the 304 heterogeneity of social associations described using the variability of the "true" SRIs estimated by a 305 Poisson model [94]. Values of S close to 0 indicate homogenous relationships within the population, 7 306 values above 0.5 reflect well differentiated social networks, and values greater than 1 indicate strong 307 differentiation. The correlation coefficient, r, between S and the observed (measured) AIs was used to 308 determine AI accuracy and their power in testing for social relationships [94].

310
Social preferences 311 To disentangle social affinities from associations that may not be driven by social preferences, we used Significance tests 322 We used permutation tests to determine the significance of observed indices of association (SRI) and 323 affiliation (GAI) for each dyadic association and across the population of whales identified on at least 5 324 occasions. The null models from these permutations were used to test the hypotheses that there were 325 strong associations and that preferences were more prevalent and stronger than expected by random 326 chance. We conducted these permutation tests for the following subsets of the humpback whale social To determine the extent to which patterns in habitat use were related to the humpback whale social 346 network, we developed separate tests for dyadic associations, the full network of associations, and the full 347 network of social preferences.

349
Dyadic associations 350 To test whether dyadic associations were behavior-specific, and whether engaging together in a certain 351 behavior increases the chances of engaging in others, we developed the following randomization routine. 352 For all dyads seen together on at least three occasions, we first determined the proportion of occasions 353 ('behavioral rates') in which bubble net feeding, other modes of feeding, traveling, resting, and robust 354 behavior were observed. We then took a behavior of interest, e.g., bubble net feeding, and identified the 355 dyads who had been encountered while engaged in that behavior ('practitioners') and the dyads that had 356 not ('others'). Between these two groups, we then compared the rate of a secondary behavior, e.g., 8 357 traveling, using a bootstrap differencing technique [108] in which random samples were drawn from the 358 two groups and the difference of those samples was recorded, and this process was repeated 10,000 times 359 to produce a distribution of differences. The mean of this distribution was used as the test statistic. We 360 compared it to a null distribution of means produced by randomizing the behavior notes in the original 361 sighting records and re-running the procedure 1,000 times. By comparing the observed mean difference to 362 the null distribution, we determined the probability that dyads engaged in the primary behavior would 363 ever be observed on a separate occasion to be engaged in the secondary behavior. If the observed mean 364 difference fell above the 97.5% quantile of the null distribution, practitioners of the behavior of interest 365 (in our example, bubble net feeding) were significantly more likely to be engaged in the secondary 366 behavior (traveling) than would be expected from random chance. If the observed mean difference fell 367 below the 2.5% quantile of the null distribution, practitioners were significantly less likely to be engaged 368 in the secondary behavior.

370
Association network 371 Our test for social niche partitioning within the association network was based upon the concept of 372 network congruency. Based on the clustering algorithm described above (see Network structure and 373 stability), we noted the community to which each individual was assigned. Then, for each possible dyad 374 pair, we noted whether or not the two individuals were assigned to the same community. We then used k-means clustering to assign individuals to communities based upon standardized numerical 388 variables that characterize some behavioral components of habitat use: mean and standard deviation of 389 fjord position (i.e., distance from the inner fjord), bubble net feeding rate (the proportion of encounters in 390 which the individual was engaged in bubble net feeding), feeding rate (referring to modes of feeding other 391 than bubble netting), social rate, and resting rate. We considered other variables (e.g., observed calving 392 rate and average minimum group size) but excluded them based on their collinearity with more 393 informative variables (e.g., observed calving rate is a function of times seen and therefore site fidelity; see 394 next stage of analysis. Also, group size is correlated to certain behaviors such as bubble net feeding, Table   395   S1). For every possible combination of these variables (n=63), we used k-means clustering in R (base 396 'stats' package; maximum iterations = 100, number of starts = 10) to assign each individual into k=7 397 communities, which is the number of communities identified in the association network structure analysis 398 described above (see Results).

400
We then calculated the congruency of these 63 behavior-based clusterings with the original social 401 clustering scheme. For each, the congruency index was then compared to the null congruency distribution 402 to determine its significance: the proportion of randomized congruency indices that was greater than the 403 behavioral congruency index was treated as a p-value. The significance (α=0.05) of a k-means clustering 404 scheme, which is based upon a certain subset of behavioral variables, indicates that exhibition of those 405 behaviors across the population is, in fact, correlated to its social structure. To double-check this social-behavioral congruency significance test, we conducted the process in reverse. 414 We built a randomization set of k-means clustering schemes (n=1,000) by shuffling each behavior 415 variable with respect to humpback whale ID (we did this only for the variable set that yielded the highest 416 congruence metric of the 63 combinations tested), then reassigning individuals to one of k=7 417 communities using the k-means algorithm. We compared these randomized clusterings to the social 418 clustering using the Adjusted Rand Index, then compared this null distribution to the realized social-419 behavioral congruency. If less than 5% of the null congruency distribution was greater than the realized 420 value, the correlation of social and behavioral clusterings was validated.

422
This k-means clustering and significance testing process was repeated for site fidelity variables 423 (proportion of years seen, mean arrival date, mean minimum stay, and Standardized Site Fidelity Index; 424 n=15 variable combinations), which were also standardized prior to clustering. Finally, the behavior and 425 site fidelity variable sets were combined and k-means clustering was repeated to determine which variable 426 combination produced the highest congruency with the social network.

428
Preference network 429 Since the Generalized Affiliation Index (GAI) is designed to isolate social preferences from the effects of 430 temporal and geographic overlap in the study area, cluster congruency tests may not be the appropriate 431 approach for testing the interaction between habitat use variables and social relationships. Instead we used 432 two analytical approaches: first, assortativity coefficient (AC) significance testing, and second, network Assortativity coefficients -We asked whether the GAI network was assorted according to tendencies in 436 behavior and/or site fidelity. To do so, we calculated ACs for each candidate variable using R package 437 'assortnet' [112]. ACs are positive if individuals with similar traits tend to positively connect, and 438 negative if strongly different individuals tend to negatively connect (i.e., avoid each other). Since both 439 GAIs and ACs can be positive or negative, we conducted this assortativity analysis separately for the 440 network of social affiliations (positive GAIs) and the network of social avoidance (negative GAIs).  (Table   480 1). These discovery metrics indicate that, although new individuals continue to be recruited, we have 481 identified the majority of humpback whales within the local population. Site fidelity 486 Interannual -The mean annual recapture rate was 50% (SD=17%, min = 37% in 2008, max = 75% in 487 2017; Table 1). Of the individuals we identified, 263 (58%) were seen in more than one year, 116 (26%) 488 were seen in 5 or more years, 49 (11%) were seen in 10 or more years. Three individuals were seen in all 489 16 years of the study (Table 2). Across the entire study, 137 individuals (30% of identified population) 490 were encountered on ≥ 10 occasions. The mean recapture rate within a season was 62% (SD=12%; Bubble net feeders -The subpopulation of known bubble net feeders (n=128 individuals; 65% of whales 505 seen 5 or more times), when compared to the remainder of the identified population, scored significantly 506 higher (p < 0.001) in all site fidelity metrics tested: annual rate of return, seasonal occupancy, 507 permanence, periodicity, Standardized Site Fidelity Index (SSFI), and the mean minimum stay. Bubble 508 netters arrived earlier than other whales (p < 0.001), but there was no significant difference in the mean 509 date of final encounter (p = 0.456), which is consistent with longer stays in the area. Based on 510 randomization tests, the LIR of bubble net feeders was significantly greater than the remainder of the 511 identified population for time lags of 1 -70 days (Fig. S3).

513
Behavior 514 Randomization tests indicated that geographic patterns of several aspects of habitat use (Fig. S4) were 515 significantly unlikely under the null hypothesis that whale behaviors were distributed randomly in space 516 and time (Fig. 3). Bubble-net feeding occurred more frequently than expected in the outer channels and 517 less frequently within inner channels; conversely, other feeding modes (including subsurface feeding  Stability of associations 539 Of the 454 whales we identified, 355 (78%) were observed in association with one another as dyads 540 (Table 3). On 5 or more occasions we observed 276 of these dyads, involving 35 individuals. 22 dyads 541 (10 whales) were observed on 10 or more occasions, and six whales were involved in 8 dyadic 542 associations that were observed at least 35 times. Many dyads (n=908 dyads of 133 whales) were 543 observed in multiple years (Table 3). Of these, 155 dyads (20 whales) were observed in at least 5 years, 544 and 17 dyads (7 whales) were observed in at least 10 years. Seven dyads (6 whales) were observed in 12 545 years.

547
Randomization tests of Lagged Association Rates indicate that, on average, dyads remained associated for 548 two months longer than would be expected based on the null model of random association-dissociation 549 ( Fig. 4; Supplementary Appendix 3). The LAR of whales occurring in groups that contained known 550 bubble net feeders (n=2,255 encounters) was significantly higher at greater time lags when compared to 551 groups that contained whales not known to bubble net feed (n=943; Fig. 4). The former LAR was 552 significantly higher than the null model up to time lags of 60 days; while the latter LAR was significant 553 within a time lag of only 5 days. Dyadic associations in which at least one individual was known to 554 bubble net feed were observed in more encounters and in more years than dyads in which neither whale 555 was known to bubble net feed (Kruskal-Wallis rank sum test, χ 2 = 192, df=1, p < 0.001).

557
Association network 558 The association network was moderately connected, realizing 14% of the total number of possible dyadic 559 associations within the identified population (Table S3). The population exhibited strong social 560 differentiation (S = 9.9, r = 0.92), providing sufficient power to test the hypothesis that KFS humpback 561 whales had no preferred or avoided relationships (S 2 x H = 9.9 2 x 11.4 >> 5; [94]). The SRI of most 562 association pairs (77%) fell between 0.01 and 0.05 (median = 0.021; mean = 0.028; SD = 0.024; Table   563 S3). The highest SRI value between any dyad was 0.438. Based on data stream permutations, strong 564 associations were more common than would be expected from random association-dissociation dynamics 565 (p < 0.001). 11% of non-zero associations were stronger than expected from chance at α=0.05.

567
Social preferences 568 We used Generalized Affiliation Indices (GAIs) to control for the influence of non-social factors (i.e., co-569 occurrence in space and time and individual gregariousness, According to data stream permutations, the prevalence and strength of social preference fell short of 574 significance (p=0.088) in tests of the entire population. However, the significance of social preferences 575 differed for subsets of the population. Within the subnetwork of known bubble net feeders (5,420 dyads 576 of 108 whales), social preferences were more common than expected (p=0.006) and stronger than 577 expected, with borderline significance (p=0.057). The strongest GAI value among bubble net feeders was 578 9.05. In contrast, for the remaining subnetwork of whales not observed to bubble net feed (1,052 dyads of 579 50 whales), social preferences were less common (p=0.800) and weaker (p=0.924) than expected, the 580 strongest GAI was lower (3.542), and a larger percentage of GAIs were negative (94%; compare to 80% 581 within the subnetwork of bubble net feeders). Social preferences were similarly weak for dyadic 582 associations between known bubble net feeders and whales not known to bubble net feed (4,867 dyads of 583 158 whales; Table S3).

585
Social network structure & stability 586 We found support for subdivision of the observed social association network into communities of The structure of the association network shifted roughly halfway through the 16-year study (Fig. S7). In 603 the early years (2004 -2011), the humpback population was dispersed into a higher number of smaller 604 communities than would be expected by chance (Fig. 6A-B), with the exception of a core network of 605 regular bubble net feeders (Fig. S7). In contrast, during the more recent years of 2012-2019, network 606 connectivity increased such that the number of discrete communities declined and the five largest 607 contained a larger proportion of the population than expected (Fig. 6A-B). During this 2010-2012 shift, community modularity broke down and was similar to what would be expected from random social 609 mixing (Fig. 6C).

611 612
Relationship between social structure and habitat use 613 614 Dyadic associations 615 Engaging together in certain behavioral contexts was correlated strongly with engaging together in others 616 (Table 4). Dyads who were seen sub-surface feeding together were more likely to travel together 617 (p=0.000) compared to those that were not, and, conversely, dyads found traveling and exhibiting robust 618 behavior were more likely to be found sub-surface feeding (p=0.004 and p=0.002, respectively). Bubble 619 net feeding together, however, did not, on average, translate to associations in other behavioral contexts. 620 Bubble net feeding dyads were significantly unlikely to be found engaged in sub-surface feeding 621 (p=0.000) or resting (p=0.004).

623
Association networks 624 Figure  traits, such as being a known mother (Fig. S9), did not appear to be assorted within the network. These Randomization tests indicated that behavior and site fidelity clusterings based on k=7 communities were 652 also significantly congruent (p < 0.001). When behavior and site fidelity variable sets were combined 653 (n=1,023 combinations), 50% of sets yielded significantly congruent (p  0.05) community assignments, 654 and 92 (9%) were congruent at p < 0.0001. Within these variable sets, the five most frequently included 655 variables were years seen (99% of variable sets), SSFI (70%), standard deviation of fjord position (68%), 656 bubble net feeding rate (59%), and social rate (56%). When mapping individual-level traits upon the network of social preferences (Fig. 7), the assortment 660 patterns described above remained evident. Based upon assortativity coefficient (AC) significance testing 661 (Table S5) for many of the traits tested, trait similarity between individuals appeared to reduce avoidance 662 and facilitate social assortment. This is indicated by lower ACs than expected within the negative GAI 663 network for the following traits: bubble net feeding rate (p=0.000), feeding rate (p=0.001), social rate 664 (p=0.001), mean fjord position (p=0.000), SD fjord position (p=0.004), and average arrival date 665 (p=0.0001). Of these, feeding rate also registered as a significant assortment factor among preferred 666 relationships (p=0.007), and bubble net feeding rate and mean fjord position fell just short of significance 667 (p=0.075 and p=0.040, respectively).

669
Interestingly, two site fidelity traits (proportion of years seen and average minimum stay) yielded 670 significantly large ACs within the negative GAI network (p=1.00 and 0.994, respectively), and the AC for 671 the categorical variable of 'known bubble net feeder' (yes or no) was nearly significant (0.900; Table S5). 672 These traits appear to be important factors for avoidance, in that whales who are dissimilar in these 673 respects tend to avoid social interaction, even when controlling for structural variables such as geographic 674 and temporal overlap.  (Table 5; see Table S8 for corresponding 678 analysis using association indices instead of affiliation indices). Whales known to bubble net feed 679 exhibited significantly higher degree-, betweenness-, and closeness-centrality and had more contacts than 680 expected (p < 0.001). Similar network position patterns were seen in whales with high annual rates of 681 return, and within-season residency (SSFI) was also significantly correlated to high centrality (p < 0.01). 682 Conversely, the position ~ trait correlation was significantly weaker than expected for whales that 683 commonly practiced other modes of feeding, and we found no correlation between network positions and 684 social rate, resting rate, or status as a known mother. Whales who were encountered primarily in offshore 685 channels exhibited high degree-and closeness-centrality (p  0.002), while whales whose fjord position 686 varied greatly exhibited high betweenness centrality (p=0.004). Whales that remained in the area for 687 longer stays exhibited higher degree centrality (p=0.003) and a higher number of contacts (p=0.001).  not an artefact of methodologies or behavior, may reflect the fact that the KFS is located more or less at 732 the geographic center of this feeding aggregation's range, and therefore may be more likely to be visited 733 while individuals transit between feeding grounds to the north and south. If this is the case, then the fjord tendencies in their choice of habitat features, prey type, feeding mode, and social group. Between these 742 two extremes, the varieties of strategies vary in four salient ways: primary mode of feeding, habitat 743 selection within the fjord system, the seasonal timing of their residency, and the variety and stability of 744 their social bonds. We found that these strategies were all correlated; for example, bubble net feeders tend 745 to use the outer channels of the fjord system, arrive early in the year, stay in the area for longer periods, 746 and exhibit stronger social preferences and greater social selectivity. Other individuals, in contrast, did 747 not engage in bubble net feeding at all, appeared to practice deep-feeding behavior (likely targeting krill; 748 [138]), occurred more commonly deeper into the fjord system, arrived later in the year, were more 749 socially fluid, but engaged in more purely 'social' behavior in the sense that their interactions were often 750 not tied in any apparent way to feeding. The fact that these social niches are manifest along a continuum 751 differs from examples of social niche partitioning observed in odontocetes such as sperm whales (e.g., 752 [45]) and killer whales (e.g., [57]), and may reflect the fact that this feeding habitat was only recently 753 reoccupied by humpback whale populations that are still in the process of recovering, in terms of both 754 abundance and ecological role, from commercial whaling [81,108]. Furthermore, it is unknown whether 755 or not the social and ecological roles manifest within this population will recover in step with its 756 abundance, if at all, nor is it known whether or not the patterns emerging in the KFS even reflect those 757 that existed before and during commercial whaling. Future years of study within this habitat will allow us 758 to interpret the stability and discreteness of the social niches we have documented here. The social niche partitioning we report here offers further insight into previous studies of habitat use 761 within this same humpback whale population and study area. Keen et al. [84] observed that humpback 762 whales occupied the KFS in a 'wave' pattern, in which whale distribution is concentrated in offshore 763 channels in the early summer, then propagates to deep inland channels in the autumn. Active acoustic 764 surveys were used to relate prey distribution to this 'whale wave', but that correlation was found to be 765 very weak, particularly in the autumn months, and the authors were ultimately unable to determine the 766 drivers of this strange habitat use pattern. That study was based on a subset of the data we have presented 767 here, but it relied only upon the locations of whale encounters, and did not include photo-identifications. 768 Here, individual sighting histories and social records have demonstrated that the 'whale wave' within this 769 fjord system is, in fact, the result of social niche partitioning. While the distribution of whales does shift 770 inland throughout the summer, we can now say that the composition of individuals is also turning over, 771 and that these individuals exhibit specific habitat preferences within the fjord system. In general, early 772 arrivals exhibit a preference for the fjord's outer channels, late arrivals prefer inner channels, and Rupert (N 54.3, unpubl. data), and as far south as Cape Caution (N 51.1, unpubl. data). As a behavior that 803 allows humpback whales to access a prey type that individuals cannot capture effectively on their own, 804 bubble net feeding is likely of critical importance to the species' ecological niche, carrying capacity, and 805 resilience to environmental perturbations within Pacific Canada [80]. In order to better understand the 806 importance of bubble net feeding to humpback whales along this coast, on both local and regional scales, 807 we emphasize the need for directed research regarding the underlying social dynamics of the behavior in 808 Canadian waters, as well as its habitat requirements, particularly with respect to its sensitivity, if any, to 809 disturbances from anthropogenic noise, fishery interactions, and nearby vessel activity. The shift in the humpback social network structure that we observed throughout the study is noteworthy 851 for its coincidence with environmental changes and demographic shifts within the same time frame. 852 Halfway through the study (2010 -2012), we saw the network structure change from a fragmented 853 community with a large number of relatively small social clusters to a more strongly interconnected 854 community with a few disproportionately large social clusters (Fig. 6, Fig. S7). This corresponds to the system, which had been increasing since we first began studying them in 2004 (Fig. S1). Connectivity 876 remained high later in the study even as encounter rates dropped, but it is conceivable that the social 877 connectivity, having been established, was preserved even as population density returned to low levels.   1  454  1  454  2  263  2  315  3  208  5  192  4  146  10  137  5  116  20  83  6  98  30  49  7  83  40  34  8  68  50  24  9  56  60  14  10  49  70  11  11  40  80  11  12  34  90  9  13  29  100  7  14  20  110  3  15  9  120  1  16  3  126  1  1402  1403  1404  Table 3 behavior would ever be observed on a separate occasion to be engaged in a secondary behavior, as determined by 1411 randomization tests (n=1,000; see main text) applied to the subset of dyads seen together on at least three occasions. 1412 Significant findings, based on a two-tailed significance test (p < 0.025 or p > 0.975) are in boldface. Column n 1413 provides the sample size of the dyad sets used in the test; the first number is the size of the dyad set known to 1414 practice the primary behavior together, and the second number is the size of the set of dyads who were not. 1415  (2.5% and 97.5% quantiles), respectively, of the permutation tests (n=100). Lags at which the running mean rises 1445 above the shaded area indicate significant patterns in residency behavior. Right: Best-fitting SOCPROG model of 1446 LIR (note log scale). Points with standard errors (n=100 bootstraps) are lag-pooled calculations of the LIR, 1447 computed at = 2 0-8 . Orange line represents the best-fitting model (see Table S6). 1448 1449