We have a paper “Recurrent seascape units identify key ecological processes along the western Antarctic Peninsula” that is now available via advance-online through the journal Global Change Biology. I place full blame for this paper on my former postdoctoral advisor and co-author Hugh Ducklow. Shortly after I arrived for my postdoc at the Lamont-Doherty Earth Observatory, Hugh suggested using all the core Palmer LTER parameters since the start of the start of that program, and “some kind of multivariate method” to identify different years. The presumption was that different years would map to some kind of recognizable ecological or climatic phenomenon.
At the time I knew nothing about seascapes or geospatial analysis. However, I had been playing around with self organizing maps (SOMs) to segment microbial community structure data. I thought that similarly segmenting geospatial data would yield an interesting result, so we gave it a go. This involved carefully QC’ing all the core Palmer LTER data since 1993 (sadly discarding several years with erroneous or missing parameters), interpolating the data for each year to build 3 dimensional maps of each parameter (you can find these data here), then classifying each point in these maps with a SOM trained on the original data. After a lot of back and forth with co-authors Maria Kavanaugh and Scott Doney, we elected to use the term “seascape unit” for different regions of the SOM. Our classification scheme ultimately maps these seascape units to the original sampling grid. By quantifying the extent of each seascape unit in each year we can attempt to identify similar years, and also identify climatic phenomena that exert controls on seascape unit abundance.
If you’re scratching your head at why it’s necessary to resort to seascape units for such an analysis it’s helpful to take a look at the training data in the standard T-S plot space.
Fig. 2 from Bowman et al., 2018. Distribution of the training data in a) silicate-nitrate space and b) T-S space. The color of the points gives the seascape unit.
The “V” distribution of points is characteristic of the western Antarctic Peninsula (WAP), and highlights the strong, dual relationship between temperature and salinity. The warmest, saltiest water is associated with upper circumpolar deepwater (UCDW) and originates from offshore in the Antarctic Circumpolar Current (ACC). The coldest water is winter water (WW), which is formed from brine rejection and heat loss during the winter. Warm, fresh water is associated with summer surface water (SW). Note, however, that multiple seascape unites are associated with each water mass. The reason for this is that nutrient concentrations can vary quite a bit within each water mass. My favorite example is WW, which we usually think of as rich in nutrients (nutrient replete); it is, but WW associated with SU 2 is a lot less nutrient rich than that associated with SU 3. Both will support a bloom, but the strength, duration, and composition of the bloom is likely to differ.
To evaluate how different climatic phenomena might influence the distribution of seascape units in different years we applied elastic-net regression as described here. This is where things got a bit frustrating. It was really difficult to build models that described a reasonable amount of the variance in seascape unit abundance. Where we could the usual suspects popped out as good predictors; October and January ice conditions play a major role in determining the ecological state of the WAP. But it’s clear that lots of other things do as well, or that the tested predictors are interacting in non-linear ways that make it very difficult to predict the occurrence of a given ecosystem state.
We did get some interesting results predicting clusters of years. Based on hierarchical clustering, the relative abundance of seascape units in a core sampling area suggests two very distinct types of years. We tested models based on combinations of time-lagged variables (monthly sea ice extent, fraction of open water, ENSO, SAM, etc.) to predict year-type, with June and October within-pack ice open water extent best predicting year-type. This fits well with our understanding of the system; fall and spring storm conditions are known to exert a control on bloom conditions the following year. In our analysis, when the areal extent of fall and spring within-pack ice open water is high (think cold but windy), chlorophyll the following summer is associated with a specific seascape (SU 1 below) that is found most frequently offshore. When the opposite conditions prevail, chlorophyll the following summer is associated with a specific seascape (SU 8) that is found most frequently inshore. Interestingly, the chlorophyll inventory isn’t that different between year-types, but the density and distribution of chlorophyll is, which presumably matters for the higher trophic levels (that analysis is somewhere on my to-do list).
Fig. 3 from Bowman et al., 2018. Clustering of the available years into different year-types. The extent of within-pack ice open water in June and October are reasonable predictors of year-type. Panel A shows the relative abundance of seascape unit for each year-type. Panel B shows the fraction of chlorophyll for each year that is associated with each seascape unit.
One of our most intriguing findings was a steady increase in the relative abundance of SU 3 over time. SU 3 is one of those seascapes associated with winter water; it’s the low nutrient winter water variant. That steady increase means that there has been a steady decrease in the bulk silicate inventory in the study area with time. I’m not sure what that means, though my guess is there’s been an increase in early season primary production which could draw down nutrients while winter water is still forming.
Fig. 5 from Bowman et al., 2018. SU 3, a low-nutrient flavor of winter water, has been increasing in relative abundance. This has driven a significant decrease in silicate concentrations in the study area during the Palmer LTER cruise.
