The Salish Sea is a semi-enclosed fjord-like estuary on the British Columbia coast, composed of the Strait of Georgia (SoG), Juan de Fuca Strait (JdF), and Puget Sound (Fig. ). The SoG is connected to the open ocean by Juan de Fuca Strait to the south and Johnstone Strait to the north, with Juan de Fuca Strait serving as the site of primary seawater exchange with the open ocean (). The Salish Sea receives freshwater input from over 200 rivers, but the primary freshwater source is the nival–glacial Fraser River (), which drives salinity-induced stratification in the CSoG and a strong estuarine exchange (). Salinity stratification is opposed by wind and tidal action. Strong winds in the fall and winter months lead to mixing of surface and intermediate water masses. The SoG contains two deep basins (north and central), with the Fraser River plume sitting on top of the central basin. Deep SoG water is relatively unmixed, except during deep-water renewal events ().

This coastal ocean is a region of ecological and cultural importance, providing habitats for important megafauna, including the southern resident killer whales (Orcinus orca) and the local salmon populations. The ongoing significant decline of the local Coho and Chinook salmon () has been implicated as a factor in the low reproductive success of the killer whale populations (), which depend on these salmon as a food source. The health of fish populations in the Pacific Northwest has been linked to spring bloom timing and phytoplankton abundance (e.g., ; ). Thus, potential population declines in upper trophic levels further motivate the understanding of factors controlling the base of the food web.

The physical environment of the Salish Sea is well-known, with functionally distinct physical–oceanographic regions (). An ongoing subject of interest in this coastal sea is the relationship between known physical and presumed ecological regions. Three prominent parts of the Salish Sea–Juan de Fuca Strait (JdF), the northern Strait of Georgia (NSoG), and the central SoG (CSoG) have been defined by distinct stratification regimes and water mass characteristics, and available biological observations and model results (e.g., ) are typically discussed in the context of these differing physical environments. However, in situ sampling of phytoplankton biomass remains relatively sparse and episodic and may not capture inherently dynamic phytoplankton biomass fluctuations, and remote sensing approaches can provide only surface chlorophyll concentrations. Here, we aim to use an unsupervised cluster analysis of a well-resolved sub-mesoscale mechanistic biophysical model to consider the linkages between the regional physical oceanography of the system and its phytoplankton biomass dynamics.

Clustering methods have demonstrated utility in identifying underlying structures in large observational datasets and are commonly used in ecological and biological observational studies. In recent years, the application of clustering to physical and biogeochemical ocean models has become more common (e.g., ), though these approaches are not yet in widespread use. The quantity of data motivates the use of clustering methods in a modeling context – even in our relatively spatially limited sub-mesoscale resolution model, 1 year of output of a single variable at hourly resolution is quite sizable (∼ 60 GB and ∼ 3×1010 individual values); the output of global circulation models is considerably larger. Well-tuned, high-resolution numerical models of complex natural systems are uniquely poised to provide insight regarding physical oceanographic regimes and overarching patterns, especially in diverse regimes where sampling efforts are sparse and often seasonally biased to fair weather. However, interpreting (even visualizing) large volumes of data poses a unique challenge; common approaches, such as monthly averaged map snapshots, may represent an oversimplification and fail to show the patterns present in the underlying system. By extracting small-data key metrics throughout the model domain, we reduce the size of the problem we are considering while keeping the key characteristics of the system that we are studying. We can then cluster these metrics in the hopes of revealing discrete dynamical regimes in complex regions. Furthermore, the clustering method is an objective classification of the system in the sense that it makes no prior assumptions about the locations of any oceanographic features that it finds.

Here our main goal is to investigate how physical dynamics in the Salish Sea objectively define regions of distinct phytoplankton biomass and functional group composition. We extract model-available proxies for four separate factors related to water column stratification: wind energy, freshwater index, water-column-averaged vertical eddy diffusivity, and halocline depth, as well as one indicator of primary productivity (depth-integrated phytoplankton biomass separated by functional group). We then cluster each factor individually in order to discuss the three major regions of the Salish Sea in the context of the spatial patterns in the yearly signals of these factors, as well as to consider their interannual variability. We finally compare spatial patterns in stratification factors to spatial patterns in phytoplankton biomass and discuss possible linkages between the two.

We use SalishSeaCast, a regional oceanographic model developed for the Salish Sea () using version 3.6 of the NEMO regional ocean modeling engine (). The physical model solves the Reynolds-averaged Navier–Stokes equations on an Arakawa-C grid with a 2 s barotropic time step, a 2 s vertical advection time step, and a 40 s baroclinic time step. Major physical model modifications since the first implementation are summarized in .

The model domain (Fig. ) is 898 (v) by 398 (u) horizontal cells with approximately 500 m horizontal resolution and 40 vertical z layers ranging from 1 m resolution at the surface to 27 m resolution at the bottom. SalishSeaCast has two open boundaries at JdF and Johnstone Strait, which are forced with eight tidal constituents and sea surface height predictions from NOAA's storm surge forecast at Neah Bay in JdF near the seaward entrance. The model is forced with over 150 rivers; the Fraser River runoff is taken from the Environment and Climate Change Canada flow gauge at Hope, BC, and the remaining rivers, as well as the Fraser River downstream of Hope, are forced by a monthly climatology (). Atmospheric forcing, including winds and solar radiation, is derived from the High Resolution Deterministic Prediction System (HRDPS), a nested 2.5 km resolution operational atmospheric model (). The HRDPS model output is too coarse to accurately resolve atmospheric conditions in the northern inlets, but its wind fields have shown good agreement with observations throughout the Strait of Georgia (). Coupled to the physical model is a NPZD-type biological model (SMELT – Salish Sea Lower Trophic Ecosystem Model; ), which is described in summary below.

The SMELT biological model represents the transfer of matter, using nitrogen as currency, between three classes of primary producers (diatoms, small flagellates, and the ciliate mixotroph M. rubrum), three classes of nutrients (nitrate, ammonia, and silicic acid), three classes of detritus (particulate and dissolved organic nitrogen, and biogenic silica), and one class of microzooplankton, with mesozooplankton grazing as a closure term. The growth rate of all three primary producer classes depends on the availability of nutrients and light and on temperature. The diatom class is assigned the highest maximum growth rate and the highest optimal light level and is the only class to take up dissolved silica – in the gleaner–opportunist framework, we consider it an opportunist class (). Small flagellates (representing phytoplankton groups such as cryptophytes) have the lowest maximum growth rate while competing better at low nitrogen levels, low light, and higher temperature. Small flagellates have the lowest minimum nutrient requirement, and we consider them the gleaner class in the gleaner–opportunist framework. The mixotroph M. rubrum has intermediate growth parameters while grazing on the flagellate class in addition to photosynthesizing. Details of phytoplankton growth rate as well as nutrient and light level preference are available in . A summary of minor updates to the model tuning since publication in is provided in Appendix B.

SalishSeaCast has been run operationally since 2014, and results from a 2013 to 2021 hindcast are available at https://salishsea.eos.ubc.ca/erddap/index.html (last access: 10 May 2022). The entire model system, including run environment, is documented at https://salishsea-meopar-docs.readthedocs.io (last access: 10 May 2022). In Appendix A we provide an evaluation of the model salinity, temperature, nitrate, dissolved silica, and chlorophyll against available observations for the years and model version analyzed, separated according to the major clusters found (Figs. –). In summary, the model shows consistently high skill across all clusters (Tables –), with Willmott skill scores for temperature and salinity ranging from 0.957–0.971 and 0.959–0.971, respectively, while comparisons with log-transformed total chlorophyll data yield scores of 0.599–0.712.

We analyzed 4 years of daily output from a hindcast of SalishSeaCast (2013–2016) using an unsupervised clustering algorithm (Ward's Euclidean distance method, see Sect. 2.3). We developed model-available year-long time series proxies (“signals”) for four different factors relating to stratification and mixing activity (wind strength, freshwater influx, vertical eddy diffusivity, and halocline depth) and one for an indicator of biological productivity (total depth-integrated biomass of three phytoplankton functional groups from the model's NPZD module). These signals were extracted for each year at each of 571 model “stations” spaced 10 model grid points apart (∼ 5 km, Fig. ). This spacing was chosen as a compromise between resolution and computing time, and we believe it represents the different regions of the Salish Sea well while being computationally manageable.

Several possible clusterings resulting from our analysis were visualized and compared for major differences (see Sect. 2.3). Results show the typical cluster structure for all 4 years for each individual factor (Sect. 3), while an example visualization of all possible clusterings of 1 year of one of the variables is available in Fig.  in Appendix C. Here, we describe the signals.

We use Ward's method (), a type of hierarchical clustering method, to cluster our data. Broadly, clustering methods are a subset of unsupervised machine learning methods used to reveal the underlying structure of a dataset by grouping similar data points. In hierarchical clustering methods, every data point is initially a single-point data cluster. At each step of the clustering, the two “closest” clusters are merged into a new cluster; this process is repeated until all points have been merged into a single cluster. Metrics of closeness vary between hierarchical clustering methods – while some methods use variations of the definition of the physical distance between clusters as a clustering criterion, Ward's method analyzes changing intracluster variance, or the “loss of information” () if they were to merge into a single cluster. In Ward's method, at each step, the clusters whose merging results in the lowest increase in intracluster variance are combined.

Many hierarchical clustering methods exist; of these we chose Ward's method because the algorithm is straightforward to implement and compares favorably to other hierarchical clustering methods with regards to performance in identifying structure in known clusterings (e.g., ). We perform hierarchical clustering using Ward's method on each of the five signals independently. For each signal, the clustering is done four times (once for each of the 4 years 2013–2016), and the results for the 4 years are then compared to assess interannual variability in the patterns found.

The physical–oceanographic dynamics of the CSoG are dominated by stratification due to Fraser River runoff, which is easily visible in the derived clustering of the freshwater index (Fig. ). Spatially, in all 4 years of our analysis, the highest freshwater index is seen near the mouth of the Fraser River and in Howe Sound (cluster 1/gold), and it then radially decreases in bands (cluster 2/grey, cluster 3/sky blue) outward from this maximum. The tendency of the surface Fraser River plume to move north from the mouth of the Fraser due to the Coriolis force () is also easily observable in this visualization. The stratifying tendency of the Fraser River (and of other major rivers) is then reflected in the clustering of the halocline signals (Fig. ). The CSoG (cluster 3/sky blue) has consistently shallow haloclines with only limited seasonal variability (∼ 5 m in summer to ∼ 7 m in winter). These shallow, stable haloclines also persist in most of the Puget Sound, owing to the influence of the Skagit River, and in the northern fjords with large rivers at their head (Toba Inlet, Bute Inlet, and Howe Sound), and the influence of these rivers is reflected in the clustering of the freshwater index. Because rivers other than the Fraser are forced by climatology in the model, the potential effects of the interannual variability of their hydrographs are not seen here.

In the wind clustering, the boundary between the CSoG and NSoG is farther south than that seen in the clusterings of freshwater index and halocline depth (Fig. ). Though winds in all clusters are highly episodic, all wind clusters show a marked decrease in wind energy during the summer months (Fig. ) – this change in mean signal magnitude and variability is most pronounced in the NSoG (cluster 4/red), which consistently shows ∼ 2 times higher wind energies in the winter months than in the summer months. In contrast, the CSoG (cluster 3/sky blue) shows the lowest variability between summer and winter energy magnitudes. Summer wind energies are actually higher in the CSoG than in the NSoG, likely due to the long wind fetch length in the CSoG, as summer winds in the Salish Sea are predominantly northerly (). Average vertical eddy diffusivity is lowest in the CSoG (Fig. ), likely owing to both high stratification and comparatively low tidal currents (), consistent with the historical idea of the Salish Sea as a system of relatively quiet basins interconnected by dynamic sills ().

As expected, the influence of the Fraser River is lower in the NSoG as the region is farther away from the river mouth (Fig. ). The resulting lower stratification is reflected in deeper and more variable haloclines in all seasons (on average, ∼ 10 m in summer to ∼ 20 m in winter) (Fig. ). A striking feature in the clustering of the freshwater index signal and halocline signals in the NSoG and the CSoG is the dissimilarity of the year 2016 to other years, reflected in a lower cluster persistence metric in this year (Fig. ). Maximum Fraser River discharge (freshet) during 2016 was remarkably low, in the lowest quartile of discharge on record, reaching only ∼ 8000 m3 s-1, or roughly 2/3 of the magnitude of the 2013–2014 freshets, which were both in the highest quartile (Fig. ). Interestingly, the mean freshwater index signal for each cluster in 2016 remains similar to the means for other years, as does the spatial extent of the most river-influenced cluster (cluster 1/gold), but the medium freshwater-influenced clusters (cluster 2/grey, cluster 3/sky blue) extend less far from the river mouth. As a result, the NSoG clusters with JdF in this year.

Dynamics in Juan de Fuca Strait are broadly characterized by limited local freshwater influence, though a small increase in freshwater index is visible in the summer months (Figs. , ), in part because of the surface advection of freshet-driven water from the CSoG due to the estuarine circulation (). The limited freshwater stratification, accompanied by a large tidal range, results in deep and variable haloclines (Fig. ). The larger tidal velocities here are also reflected in slightly higher water-column-averaged VED (Fig. ). Interestingly, in 2015, the VED in much of JdF clusters with the SoG, possibly due to the inhibition of water column mixing by higher thermal stratification of the system due to the significant marine heatwave in the North Pacific in the years 2013–2015 (), whose effects were most pronounced in the Salish Sea in 2015 (). The dissimilarity of the year 2015 to other years is reflected in the cluster persistence metric (Fig. ).

Water-column-averaged vertical eddy diffusivity in the Salish Sea is dominated by tidal mixing activity (), allowing clustering VED to uncover dominant tidal hotspots. VED varies by 3 orders of magnitude in the model domain (Fig. ). As expected, this metric reaches its maximum in the Haro Strait region, as well as in parts of Puget Sound, for example in Admiralty Inlet, near known tidal mixing hotspots (). Two stations in northern Johnstone Strait and the Discovery Passage region also exhibit heightened VED in all 4 years, consistent with the high observed tidal velocities near Seymour Narrows in this region (). Fourier analysis of the annual vertical eddy diffusivity signals also shows local maxima in energy at weekly and fortnightly frequency (not shown) in all 4 years in all five clusters, consistent with the role of tides as the dominant source of mixing energy in the system ().

The by-cluster seasonally averaged means and standard deviations of average VED are consistent interannually (Figs. , ). The same three stations in the San Juan islands (cluster 4/red) report the highest VED in all 4 analyzed years, exhibiting maxima that are almost a factor of 2 larger than the next-largest signal (cluster 2/grey). In the highly variable Haro Strait and Johnstone Strait regions, the spatial frequency of our sampling likely plays a role in our derived map of tidal mixing hotspots – as we sample only approximately every 100th horizontal model coordinate, we likely miss other high-VED model points in this subregion, especially channels that have width scales comparable to our model resolution (0.5 km), for example the intricate channel passages of the San Juan and Discovery Island groups in the Haro and Johnstone Strait regions, which are known tidal mixing hotspots (Fig. ). Analysis of tidal mixing hotspots is not the focus of this work, but a full characterization of this tidally mixed zone using a more refined clustering approach may be an interesting focus of future work.

A similar biological clustering arises in all 4 years (Figs. , ). The boundaries of this clustering broadly coincide with the three major oceanographic subregions discussed above. The largest cluster (the CSoG – cluster 3/sky blue) is characterized by diatoms blooming first, followed by a transition to flagellate abundance in the summer months. In all 4 years, a functionally distinct NSoG region (cluster 4/red) arises, with sharp, episodic spikes in summer diatom biomass and diminished flagellate biomass. JdF (cluster 5/dark blue) reaches maximum biomass later in the year and, like the NSoG, shows a persistence of summer diatoms and diminished summer flagellate biomass. In contrast to the NSoG, where diatom biomass diminishes between episodic spikes, diatom biomass in JdF typically remains above 20 mmol N m-2 throughout the spring and summer seasons, with occasional spikes to higher biomass.

These three main regions have roughly similar mean seasonal biomass, with interannual variability larger than variability between clusters; the main differences between them are in the relative abundances of different functional groups and in the temporal characteristics of the phytoplankton biomass. Nearshore areas cluster together (cluster 2/grey) and have low depth-integrated biomass because they are limited by shallow depth. The largest depth-integrated biomass in the model in both the diatom and flagellate groups is found in the tidal mixing region of Haro Strait (cluster 1/gold).

In the model, the NSoG shows only slightly higher depth-averaged phytoplankton biomass in all seasons than the CSoG (Fig. ). This biomass is consistent with the in situ study of Masson and Peña conducted between 2001 and 2007 in this region, which shows lower surface chlorophyll but a deeper phytoplankton growing zone in the NSoG, leading to slightly higher depth-integrated chlorophyll concentrations in the northern region in all four seasons of sampling (, henceforth MP09, their Table 2). Remote sensing observations also show significantly lower surface chlorophyll in the NSoG throughout the year, as well as finding anomalously high surface chlorophyll concentrations in 2015 that are not reproduced by the model (). The majority of the modeled biomass difference between the NSoG and CSoG occurs in the subsurface maximum around 6–8 m in depth (Fig. ), where it cannot be detected by remote sensing. Previous modeling (of the years 2007–2009) found somewhat higher depth-integrated phytoplankton biomass in the CSoG throughout the year, but with significant spatiotemporal variability (). However, a recent year-round in situ campaign in several parts of the Strait of Georgia found no meaningful difference in depth-integrated chlorophyll between the NSoG and CSoG ().

Together, these studies suggest that the difference between the two regions with respect to total depth-integrated biomass is subtle. However, we find a substantial difference in the modeled phytoplankton functional group composition and the temporal scale of variability of the phytoplankton signal between regions. In both regions of the strait, the opportunist-class diatoms, which have the highest growth rate and highest nutrient requirements, peak first (typically in late March, though with considerable variability; ) and form the majority of the phytoplankton biomass in the spring (Figs. , ). In the CSoG, the model then transitions to higher biomass of gleaner-type flagellates around day 150, near the beginning of June, and flagellates continue to exhibit high summer biomass in this region (Figs. , ). In contrast, the NSoG continues to exhibit episodic short-lived peaks of high opportunist-type phytoplankton biomass, represented by diatoms, throughout the summer so that in all years except 2016, diatoms make up the majority (∼ 55 %–60 %) of summer phytoplankton biomass in this region.

In the Strait of Georgia, significant evidence of high summer biomass near strong mixing zones or in response to mixing driven-nutrient delivery exists. For example, early surveys of the system find high chlorophyll associated with dynamic frontal regions in the northern and southern ends of the SoG (), and they consequently warn against drawing firm conclusions about the nature of phytoplankton abundance and variability from episodic sampling in shifting frontal zones. Nutrient delivery via episodic tidal mixing events near Discovery Passage has been linked to increased biomass (e.g., ) and modeled primary productivity (). Though this phenomenon has been recorded in the CSoG as well (), higher stratification may dampen the magnitude of the nutrient pulses. Sudden introduction of abundant nutrients is expected to favor the opportunist functional group represented by diatoms over the slower-growing gleaner functional group represented by flagellates, as is seen in our clustering (). A recent 4-year time series of phytoplankton composition data at a station near Quadra Island in the NSoG supports this idea by showing episodic blooms of summer diatoms after wind events (). Indirect evidence of episodic high biomass, sometimes following wind events, has been observed elsewhere in the NSoG ().

We suggest that our results reflect a controlling influence of stratification on phytoplankton biomass and community structure. Strong stratification concentrates phytoplankton biomass in a thin well-lit surface layer while limiting supply of nutrients after the initial biological drawdown. In the model, these conditions favor high abundance of the gleaner flagellate group. In the NSoG, nutrient drawdown also occurs, but episodic wind events lead to stronger upwelling and mixing due to the comparatively weaker stratification and inject sharp pulses of nutrients into the near surface, leading to sharp, short-lived diatom blooms (). In contrast, in the CSoG, despite stronger summer winds, strong stratification continues to favor gleaner-type organisms. Faster-growing opportunist diatoms tend to outcompete gleaner flagellates when sufficient nutrients and light are available, but inherent variability in the physical environment promotes coexistence (). The result is only a modest, if any, change in biomass but a significant change in functional group composition and temporal variability between the NSoG and CSoG.

Peña et al. find higher biomass in the CSoG due to the deeper nutricline in the NSoG (). We find instead that the increased mixing in the NSoG provides increased nutrients and that biomass in both regions is about the same. These two views are not directly reconcilable and which view is more representative of actual conditions depends on accurately capturing the balance between the action of mixing as a source of nutrients and mixing as a source of light reduction and phytoplankton dilution.

The extent of phytoplankton diversity in the Salish Sea cannot be strictly condensed into the three functional groups represented in this model. The divide of the phytoplankton functional groups in the model does not precisely correspond to a split between diatoms and all types of flagellates. For instance, silicoflagellates (class Dictyochophyceae) might align with the diatom class based on silicon utilization. For this reason, we discuss these classes in terms of competition between opportunist-type primary producers with high nutrient needs and high light needs and gleaner-type primary producers with capacity to persist at lower nutrient and light levels. For this study, the requirement is capturing the overall regional biomass patterns and function. Here we provide a brief comparison between our results and available in situ phytoplankton functional group observations.

In situ measurements of the relative abundance of phytoplankton functional groups remain rare in the Salish Sea and tend to be sparse in space and/or time. In situ observations represent a snapshot at a single station and depth, while model output instead presents the average of a much larger volume (the discrete model cell). A recent temporally well-resolved 4-year (2015–2018) time series of phytoplankton biomass and composition, derived from high-performance liquid chromatography (HPLC) analysis of phytoplankton pigments, taken at a single station in the NSoG () provides a starting point for such a comparison. The in situ data, taken from a depth of 5 m, show diatom-dominated blooms with varying start dates in the spring season, followed by a transition to a regime wherein flagellate-type groups (chiefly prasinophytes and cryptophytes) make up the majority of phytoplankton biomass, but diatoms remain present (, their Figs. 4, 5). Episodic later-summer diatom blooms occur in 3 of the 4 observational years, corroborating the modeled later-summer NSoG diatom blooms seen in this study.

Local phytoplankton composition data derived from shipboard observations from spring, summer, and autumn cruises spanning the Juan de Fuca Strait and both the CSoG and the NSoG are also available as a technical report by . Unlike the Del Bel Belluz study, these data are relatively well-resolved in space but less resolved temporally; in a given year, typically only 1 d of observations is available for each station. The relative abundance of phytoplankton functional groups in the three regions is thus interannually variable (, their Fig. 39-2), and in contrast to the Del Bel Belluz study, the data show only limited summer presence of diatoms in the NSoG in either of the years overlapping with our study (2015 and 2016). These observations contradict trends seen in our model. Simultaneously, these data show summer diatom dominance in the well-mixed Haro Strait region, corresponding to our tidal mixing region, which echoes trends we see in this study. Taken together, these two in situ phytoplankton composition studies each provide some corroboration of patterns we see in the model but differ from each other regarding summertime diatom representation in the NSoG. Combined with the modeling, these three perspectives on phytoplankton composition and biogeography each represent different spatial and temporal scales. The questions raised by their contrasting findings highlight the need for both modeling and observational work to provide a holistic view of the local biophysical dynamics.

Our modeling study is necessarily subject to limitations. For example, very high biomass shown in the tidal mixing region (Fig. , cluster 1/gold) could be an artifact of slower phytoplankton mortality rates, at least at times, than occur in nature, with phytoplankton mixed deep into the water column and persisting too long. Such a rate imbalance would affect the response to mixing described above. Available observations support model phytoplankton levels in this region but are limited to the upper water column. Because model–data agreement on biomass and nitrate is strong in these regions, we believe the mechanism of nutrient delivery by wind events in the less stratified north, leading to dominance of faster-growing phytoplankton, is robust.

Our results suggest that the mean annual average depth-integrated biomass is about the same in all three physical regions, including the well-mixed, weakly stratified JdF. In contrast, previous studies suggested a lower biomass in JdF () due to a deeper nutricline. However, recent in situ chlorophyll and nutrient data (2013–2016) support our result. In fact, the evaluation suggests that, for dates and locations for which observations are available, the model slightly underestimates observed biomass in Juan de Fuca Strait (Fig. , Table ).

One factor contributing to the difference between these conclusions may be the vertical structure in the biomass observed by both MP09 and the model. In MP09, the spring phytoplankton biomass is much more prominent in the CSoG and NSoG than in JdF. The spring biomass exhibits a strong subsurface maximum (∼ 10 m in the chlorophyll observations) and persists relatively deep into the water column (up to 40 m). However, though overall concentrations reported in MP09 are lower in all seasons in JdF, observed chlorophyll concentrations >= 1 mg m-3 persist at deeper depths in most seasons in Juan de Fuca than in both regions of the Strait of Georgia (up to 50 m in the spring, summer, and fall), and in summer and fall, the NSoG exhibits slightly deeper phytoplankton persistence compared to the CSoG. We replicate these trends in general vertical structure (Fig. ), with a prominent subsurface maximum at ∼ 6–8 m and phytoplankton biomass mixed deeper in Juan de Fuca Strait than in either region of the Strait of Georgia.

This vertical structure likely leads to a dilution effect – even when phytoplankton concentration at a given depth may be lower in Juan de Fuca Strait than in the Strait of Georgia, overall depth-integrated biomass may be simultaneously higher. This deep biomass is less likely to be captured by sampling campaigns, potentially leading to an underestimation of the phytoplankton biomass of the region as a whole. Furthermore, because of the interannual variability in spring bloom timing and differences in spring bloom timing between the Strait of Georgia and Juan de Fuca Strait (discussed further below), the spring in situ survey that captures high biomass in the Strait of Georgia may be too early to observe the full extent of the spring bloom in Juan de Fuca Strait.

In the NSoG and CSoG, the derived biological signals suggest a regime wherein stable growing conditions in the spring transition to varying degrees of summer nutrient limitation which are interrupted by episodic nutrient delivery more frequently in the NSoG. In contrast, the diatom growth curve in JdF suggests a light-limited environment year-round, consistent with the established understanding (). Nutrients are rarely limiting in JdF owing to the plentiful supply of oceanic nitrate () and stronger water column mixing in this region demonstrated in the VED clustering (Fig. ). One factor that may potentially enhance growing conditions in the summer season here is the advection of a freshwater lens from the Strait of Georgia via the surface estuarine circulation (). This advection is visible as a slight increase in the summer freshwater index in JdF (Fig. ); it may contribute to increased water column stability and hence light availability and favorable growing conditions in this period.

Our clustering approach identifies unambiguous regions of a complex coastal sea that exhibit distinct biological responses to disparate physical environments. These responses are not immediately obvious in time-averaged snapshots of the studied system. The simple machine learning technique used here enhances our way of looking at the problem – in this application, we are not using machine learning to predict unknown quantities, as is becoming common (e.g., ), but instead we are asking it to show us what is already there. Using this simple technique, we are able to draw objective boundaries between regions based on emergent structures in our data and significantly advance our intuition about the system. Cluster-based model evaluation may also be a very useful application of clustering techniques, as it has potential to diagnose how well a given model performs across different biophysical regimes.

The simplicity of the approach may have utility in numerous contexts. For instance, many characterizations of environmental regions rely on sparse data with large spatial biases. Objective clusters determined from regional models, with mechanistic underpinnings, may be used to group sparse data. This approach allows clear characterization of complex systems. Furthermore, it may provide the necessary first step for machine learning studies that rely on well-organized training datasets to accurately predict target variables (e.g., ). Resource and environmental management situations and optimal monitoring strategies may also benefit from a data-driven approach to regional definitions.