Comparing mechanistic and empirical model projections of crop suitability and productivity: implications for ecological forecasting
Article first published online: 7 JAN 2013
© 2013 John Wiley & Sons Ltd
Global Ecology and Biogeography
Volume 22, Issue 8, pages 1007–1018, August 2013
How to Cite
Estes, L. D., Bradley, B. A., Beukes, H., Hole, D. G., Lau, M., Oppenheimer, M. G., Schulze, R., Tadross, M. A. and Turner, W. R. (2013), Comparing mechanistic and empirical model projections of crop suitability and productivity: implications for ecological forecasting. Global Ecology and Biogeography, 22: 1007–1018. doi: 10.1111/geb.12034
Editor: Niklaus Zimmermann
- Issue published online: 3 JUL 2013
- Article first published online: 7 JAN 2013
- Princeton Environmental Institute's Grand Challenges Program
- ecological niche model;
- empirical model;
- habitat suitability model;
- mechanistic model;
- South Africa;
- species distribution model
Intercomparison of mechanistic and empirical models is an important step towards improving projections of potential species distribution and abundance. We aim to compare suitability and productivity estimates for a well-understood crop species to evaluate the strengths and weaknesses of mechanistic versus empirical modelling.
We compared four habitat suitability models for dryland maize based on climate and soil predictors. Two were created using maximum entropy (MAXENT), the first based on national crop distribution points and the second based only on locations with high productivity. The third approach used a generalized additive model (GAM) trained with continuous productivity data derived from the satellite normalized difference vegetation index (NDVI). The fourth model was a mechanistic crop growth model (DSSAT) made spatially explicit. We tested model accuracy by comparing the results with observed productivity derived from MODIS NDVI and with observed suitability based on the current spatial distribution of maize crop fields.
The GAM and DSSAT results were linearly correlated to NDVI-measured yield (R2 = 0.75 and 0.37, respectively). MAXENT suitability values were not linearly related to yield (R2 = 0.08); however, a MAXENT model based on occurrences of high-productivity maize was linearly related to yield (R2 = 0.62). All models produced crop suitability maps of similarly good accuracy (Kappa = 0.73–75).
These findings suggest that empirical models can achieve the same or better accuracy as mechanistic models for predicting both suitability (i.e. species range) and productivity (i.e. species abundance). While MAXENT could not predict productivity across the species range when trained on all occurrences, it could when trained with a high-productivity subset, suggesting that ecological niche models can be adjusted to better correlate with species abundance.