Predicting Species Distributions in California’s Central Valley

The development of sophisticated species distribution modeling techniques provides an opportunity to greatly improve the value of species occurrence data. Using statistical and machine learning approaches, point occurrence data can be related to GIS-based environmental data layers to generate robust, spatially continuous predictions of species occurrence. Species distribution models can be as specific or general as their data inputs. Using basic climate variables such as temperature and precipitation, general species envelopes can be developed. With the inclusion of reliable vegetation and landcover data, much more refined, habitat-based distributions may be generated.

Our approach was to use a machine learning algorithm called MaxEnt (Phillips et al. 2006), which has consistently outperformed other distribution modeling techniques, to predict species distributions based on two different datasets: (1) the California Natural Diversity Database (CNDDB); and (2) PRBO’s long-term avian point count survey data. We used the CNDDB data to model many of the California bird species of special concern (BSSC) that are not well-captured by point count surveys (i.e., non-passerine species). For other species of interest—California Partners in Flight (CPIF) bird conservation plan (BCP) focal species—we primarily used PRBO point count data to derive species’ occurrence locations.

Species distribution models were developed using the following inputs: (1) point-based species occurrence data; and (2) GIS-based environmental data layers for California (100-m x 100-m pixel resolution). A variety of vegetation, climate, hydrology, and land use data layers were manipulated to create input data layers of hypothesized importance for the species of interest (see Table 3). Manipulation of input data was performed using ESRI ArcGIS 9.2 and Fragstats 3.3 (McGarigal and Marks 1995). NDDB records were filtered by spatial accuracy and by season (only breeding season records were used). PRBO records, which have high spatial accuracy, were filtered by season only.

Predicted distributions generated by MaxEnt represent cumulative percent of occurrences. That is, at a given location, the value of that pixel represents the percent of that species’ distribution that is contained within all of the pixels of equal or lower value. The higher the value, the higher the probability of a species’ occurrence.

For more information about these models please contact Dennis Jongsomjit or Diana Stralberg
PRBO Conservation Science (www.prbo.org)

References:
McGarigal, K., and B. J. Marks. 1995. FRAGSTATS: spatial pattern analysis program for quantifying landscape structure. USDA For. Serv. Gen. Tech. Rep. PNW-351. [URL: http://www.umass.edu/landeco/research/fragstats/fragstats.html].

Phillips, S. J., R. P. Anderson, and R. E. Schapire. 2006. Maximum entropy modeling of species geographic distributions. Ecological Applications 190:231-259.