Assessing ecological niche models for species occurrence data with biases and/or errors

Background/Question/Methods

Open-access databases can store an enormous amount of information about species occurrences, which should be leveraged for analyzing global patterns of biodiversity. However, to do this effectively we must overcome problems derived from both the biases and errors contained in these databases. We tested which ecological niche modelling methods produce better predictions when calibrated with data samples that have i) biases, and ii) biases and errors. We tested two different methods, one complex, Maxent, and one simple, Bioclim. We created a virtual species, sampled its distribution with biases and errors, and calibrated the models with those samples.

Results/Conclusions

Results indicate that Bioclim produces better predictions than Maxent when calibrated with biased data sets. Bioclim does not overestimate the species’ range and is able to produce accurate predictions even when calibrated with small and biased data samples (25-50 points). However, when incorrect occurrences were included in the calibration samples, Bioclim over-predicted the species’ range. Our experiments indicate that in that case, Maxent predictions remain robust and provide accurate maps. Thus, if calibration data samples contain only biases, Bioclim provides better maps than Maxent. However, when samples include both biases and incorrect occurrences, Maxent models provide better results than Bioclim.