Modeling Methods and Conservation
NCEAS model comparison project
I am actively participating in a NCEAS working group that is comparing the performance of species distribution (ecological niche) modeling methods, using methods that require only species presence (as opposed to presence/absence or abundance). We are using data from six different regions, each of which contains: ad hoc (i.e., museum, herbarium) presence-only species locality data and high quality (i.e., environmentally and geographically stratified sampling) presence-absence species locality data; and environmental grids. We are using the presence-only (and in some cases pseudo-absence) and environmental grid datasets to develop models with a range of available modeling methods, including: BIOCLIM, DOMAIN (presence-only boxcar models), maximum entropy, GARP (genetic algorithms for range prediction), general linear models, general additive models, multiple additive regression splines, generalized dissimilarity models and boosted decision trees. The independent presence-absence data are used to evaluate predictive success/accuracy. We have developed a series of experiments related to sample size, bias, and error of point localities and grain size of environmental layers to test model performance across a variety of scenarios.
Modeling methods for rare species, climate change prediction, and alternative data sources (remote-sensing)
I am involved in a series of projects focused on modeling methods, data sources and conservation decision making. First, we are using Natureserve’s heritage database to explore how different modeling methods (BIOCLIM, Maxent, DOMAIN and GARP) perform with low sample sizes. Based on this knowledge we hope to improve predictive capacity and biological understanding of rare species. Second, we are evaluating the performance of ecological niche models for use in future climate change modeling. Our approach focuses on comparing physiological based models of how species should respond to climate changes with ecological niche models. Third, we are exploring how the use of different data sources and modeling decisions influence maps of species richness.
In a final project, we are exploring the use of remote-sensing data for predictive distributional modeling across a variety of species and geographical scenarios. This is particularly challenging because species localities, which are often from museums, have geographic uncertainty that is far greater than many fine scale remote-sensing imagery. Hence, it is difficult to use this imagery to produce fine scale predictions of species distributions. To resolve this we are partitioning species locality data based on locational accuracy and using accurate points with high resolution data layers (i.e., remotely-sensed images) and all points with low resolution data (i.e., extrapolated climate surfaces). In this way we can combine different types and qualities of data to produce species distributional models.