Both real-time and long-term environmental data perform well in predicting shorebird distributions in managed habitat

Highly mobile species, such as migratory birds, respond to seasonal and
inter-annual variability in resource availability by moving to better
habitats. Despite the recognized importance of resource thresholds,
species distribution models typically rely on long-term average habitat
conditions, mostly because large-extent, temporally-resolved,
environmental data are difficult to obtain. Recent advances in remote
sensing make it possible to incorporate more frequent measurements of
changing landscapes; however, there is often a cost in terms of model
building and processing and the added value of such efforts is unknown.
Our study tests whether incorporating real-time environmental data
increases the predictive ability of distribution models, relative to using
long-term average data. We developed and compared distribution models for
shorebirds in California’s Central Valley based on high temporal
resolution (every 16-days), and 17-year long-term average, surface water
data. Using abundance-weighted boosted regression trees, we modeled
monthly shorebird occurrence as a function of surface water availability,
crop type, wetland type, road density, temperature, and bird data source.
While modeling with both real-time and long-term average data provided
good fit to withheld validation data (0.79 < AUC < 0.89
across taxa), there were small differences in model performance. The best
models incorporated long-term average conditions and spatial pattern
information for real-time flooding (e.g. perimeter-area ratio of real-time
water bodies). There was not a substantial difference in the performance
of real-time and long-term average data models within time periods when
real-time surface water differed substantially from the long-term average
(specifically during drought years 2013-2016) and in intermittently
flooded months or locations. Spatial predictions resulting from the models
differed most in the southern region of the study area where there is
lower water availability, fewer birds, and lower sampling density.
Prediction uncertainty in the southern region of the study area highlights
the need for increased sampling in this area. Because both sets of data
performed similarly, the choice of which data to use may depend on the
management context. Real-time data may ultimately be best for guiding
dynamic, adaptive conservation actions whereas models based on long-term
averages may be more helpful for guiding permanent wetland protection and
restoration. --