Final Technical Report: - Southwest Fisheries Science Center - NOAA

of these choices as possible and used the choices that resulted in the best predictive models. To
evaluate predictive power, we used cross-validation (leaving out one survey year and predicting
densities for that year with models built using only the other years). Data from the two most
recent surveys (2005 in the CCE and 2006 in the ETP) were used for this model validation step.
We explored three modeling approaches to predict cetacean densities from habitat
variables: Generalized Linear Models (GLMs) with polynomials, Generalized Additive Models
(GAMs) with nonparametric smoothing functions, and Regression Trees. Within the category of
GAMs, we tested and compared several software implementations. In summary, we found that
Regression Trees could not deal effectively with the large number of transect segments
containing zero sightings. GLMs and GAMs both performed well and differences between the
models built using these methods were typically small. Different GAM implementations also
gave similar, but not identical results. We chose the GAM framework to build our best-and-final
models. In some cases, only the linear terms were selected, making them equivalent to GLMs.
We explored the effects of two aspects of sampling scale (resolution and extent) on our
cetacean density models. To explore the effect of resolution, we sampled transect segments on
scales ranging from 2 to 120 km. We found that differences in segment lengths within this range
had virtually no effect on our models in the ETP, but that scale affected the models for some
species in the CCE where habitats are more geographically variable. For our best-and-final
models, we accommodated this regional scale difference by using a longer segment length in the
ETP (10 km) than in the CCE (5 km). To explore the effect of extent, we constructed models
using data from the ETP and CCE separately and for the two ecosystems combined. We found
that the best predictive models were based on data from only one ecosystem; therefore, all our
best-and-final models are specific to either the CCE or the ETP.
We explored five methods of interpolating oceanographic measurements to obtain
continuous grids of our in situ oceanographic habitat variables. Cross-validation of the
interpolations gave similar results for all methods. Ordinary kriging was chosen as our preferred
method because it is widely used and because, qualitatively, it did not produce unrealistic “bull’s
eyes” in the continuous grids.
We explored the use of CCE oceanographic habitat data from two available sources: in
situ measurements collected during cetacean surveys and remotely sensed measurements from
satellites. Only sea surface temperature (SST) and measures of its variance were available from
remotely sensed sources, whereas the in situ measurements also included sea surface salinity,
surface chlorophyll and vertical properties of the water-column. We conducted a comparison of
the predictive ability of models built using in situ, remotely sensed, or combined data and found
that the combined models typically resulted in the best density predictions for a novel year of
data. In our best-and-final CCE models we therefore used the combination of in situ and
remotely sensed data that gave the best predictive power.
xvi