Testing unidimensional species distribution models to forecast and hindcast changes in marsh vegetation over 40 years

Abstract

• We test the performance of projecting over time of unidimensional SDMs. • We compare the performance of modelling methods and data types in prediction over time. • There is no significant difference between internal and external evaluation. • Cover is a little better than P/A model, and GAM is similar with eHOF model. • Our result provide guideline for SDMs in prediction under environmental change. Species distribution models (SDM) predicting changes in species occurrences and abundance are increasingly being used as a tool in biogeography and conservation biology. However, we have little information on their predictive performance. Here we used archive-recorded predictor and field-observational verifier data associated with water level to evaluate the performance of response curves over 40 years for marsh plant species in Northeast China. A consensus approach (AUC: area-under-curve) was used as the test measure for internal evaluation and external evaluation (forecast and hindcast). Our results demonstrated that there is no significant differences between internal and external evaluation, and they both showed reasonable accuracy (AUC = 0.73, respectively). There was considerable variation across species and projection direction in model accuracy, and accuracy of model fitting in internal evaluation and restricting the environmental range of data in different time periods may impact the performance of model projection over time. The performance of generalized additive models (GAM) is similar with that of extended Huisman-Olff-Fresco models (eHOF). Cover model is a little better than presence/absence models in prediction over time. Our findings provide some guidelines for the use of SDM for predictions under environmental change.