Evaluating species distribution models with field experiments: A test-case with eastern hemlock (Tsuga canadensis)

Background/Question/Methods

Models that relate species presence/absence data to environmental variables are commonly used to predict habitat suitability and species distributions. While a substantial body of literature exists on model evaluation in silico, few attempts have been made to evaluate models using actual field experiments. We constructed two species distribution models for eastern hemlock (Tsuga canadensis) using maximum entropy and generalized linear models. We were particularly interested in our models’ predictions of numerous apparently suitable, but unoccupied, habitat patches in central Indiana where hemlock has a sporadic distribution. To evaluate the model predictions we conducted field experiments using a full factorial design with two variables: presence or absence of hemlocks, and high or low quality habitat as defined by our models. 240 hemlock seedlings were planted in 12 plots in 2011 and 720 seedlings were planted in 36 plots in 2012. Seedling survivorship and soil moisture content were recorded at two-week intervals during the growing season. Data loggers were used in 2012 to collect data on temperature and light intensity at 15 minute intervals for the duration of the experiment.

Results/Conclusions

Seedlings planted in high quality habitat plots survived at significantly greater rates for the first six weeks in 2011 and the first eight weeks in 2012 (25.6% survivorship increase in 2011 and 22.1% increase in 2012, P values <0.01). In subsequent weeks, seedlings in high quality habitat maintained a higher survival rate, though not at significant levels (P values from 0.12 to 0.46), likely due to the severe drought conditions experienced mid-summer during both field seasons. Hemlock presence had no significant effect on survivorship during pre-drought conditions, but significantly decreased seedling survival in drought conditions (-12.5% in 2011, -5.7% in 2012). Hemlock presence decreased daily average light levels (P<0.01) but habitat quality had no significant effect (P=0.29). While daily average and daily minimum temperatures did not vary significantly between treatments, both hemlock presence and good quality habitat resulted in lower daily maximum temperatures (P<0.01). Soil moisture levels were consistently lower if hemlock was present or if the habitat was classified as good. In summary, seedlings survived at higher rates in good quality habitat regardless of hemlock presence, indicating that our species distribution models accurately predict areas of high quality habitat that are currently unoccupied.