5. Conclusions

This systematic review has shown that a variety of machine learning (ML) methods has been used in species threat and conservation studies. The flexibility, accuracy and growing potential of ML is now unavoidable in conservation science, although some methodological advances are still scarcely used. Often, this is due to the challenging interpretability of black-box models, reducing their value when interpretation is more important than prediction. This may be one of the reasons for the still scarce use of deep learning in the area, despite the well-known advantages in the performance of such models. The other reason may be the need for large amounts of data to train deep learning models.
Maximum entropy (MaxEnt) and Bayesian methodologies are popular ML methods in conservation, being both frequently implemented and powerful. Previously in conservation, there was a large focus on local and experimental settings, which pushed the widespread adoption of Bayesian statistics. Over the years, focus in conservation science has partially shifted to larger and more global problems, with ecosystem trends being seen as a natural extension of global trends. Owing to this fact, models like MaxEnt, which can make efficient use of spatial data, rose in popularity and seem inescapable. MaxEnt has the advantage of easily dealing with data of different types, discrete, continuous and Boolean, which, applied over geospatial data, makes most environmental features manageable and available for incorporation in large-scale models.
ML is a field still being actively developed. Existing methods are not consistently used for singular tasks, with some being almost entirely interchangeable. As such, compared with other fields, like statistics, usage of ML in conservation is defined more by accuracy and results, less by established protocols. Further studies analysing the strengths and weaknesses of each ML method in contrasting conservation problems should be encouraged.
We have shown that ML has the potential to empower, inform and simplify crucial conservation work. Its use should be encouraged among both established conservationists and students; however, only with an understanding of the field as a whole can meaningful work be done. To better succeed as conservationists, we must keep track of new developments in ML, since they often can produce qualitative improvements on previous results.

Acknowledgments

VVB and PC have both been funded by Kone Foundation grants (Koneen Säätiö), Finland. LC was partially funded by FCT through the LASIGE Research Unit, ref. UIDB/00408/2020 and ref. UIDP/00408/2020, Portugal.