Jump to Main Content
Evolutionary algorithms for species distribution modelling: A review in the context of machine learning
- Gobeyn, Sacha, Mouton, Ans M., Cord, Anna F., Kaim, Andrea, Volk, Martin, Goethals, Peter L.M.
- Ecological modelling 2019 v.392 pp. 179-195
- algorithms, artificial intelligence, biogeography, calibration, decision making, ecosystems, guidelines, models
- Scientists and decision-makers need tools that can assess which specific pressures lead to ecosystem deterioration, and which measures could reduce these pressures and/or limit their effects. In this context, species distribution models are tools that can be used to help asses these pressures. Evolutionary algorithms represent a collection of promising techniques, inspired by concepts observed in natural evolution, to support the development of species distribution models. They are suited to solve non-trivial tasks, such as the calibration of parameter-rich models, the reduction of model complexity by feature selection and/or the optimization of hyperparameters of other machine learning algorithms. Although widely used in other scientific domains, the full potential of evolutionary algorithms has yet to be explored for applied ecological research. In this synthesis, we study the role of evolutionary algorithms as a machine learning technique to develop the next generation of species distribution models. To do so, we review available methods for species distribution modelling and synthesize literature using evolutionary algorithms. In addition, we discuss specific advantages and weaknesses of evolutionary algorithms and present a guideline for their application. We find that evolutionary algorithms are increasingly used to solve specific and challenging problems. Their flexibility, adaptability and transferability in addition to their capacity to find adequate solutions to complex, non-linear problems are considered as main strengths, especially for species distribution models with a large degree of complexity. The need for programming and modelling skills can be considered as a drawback for novice modellers. In addition, setting values for hyperparameters is a challenge. Future ecological research should focus on exploring the potential of evolutionary algorithms that combine multiple tasks in one learning cycle. In addition, studies should focus on the use of novel machine learning schemes (e.g. automated hyperparameter optimization) to apply evolutionary algorithms, preferably in the context of open science. This way, ecologists and model developers can achieve an adaptable and flexible framework for developing tools useful for decision management.