3.
Results
A part of the selected papers used methods that were not explicitly
searched for, such as Apriori, multivariate adaptive regression splines
(MARS), discriminatory analysis (DA), clustering models (such ask -means clustering, hierarchical clustering analysis),
forecasting models (such as conditional autoregressive models – CAR),
and kernel density estimation (KDE). Other papers used methods that are
more traditionally associated with traditional statistics, despite
commonalities with machine learning methods, such as linear regression
(LR), generalised linear models (GLM), generalised additive models (GAM)
and mixed models, both linear and additive (GLMM, GAMM). None were
included in our analyses.
3.1 Threats
The methods most used for threat analysis were MaxEnt, Bayesian and
Boosting (Fig. 1, Annex D). The most studied threat according to the CBD
classification was Habitat loss & fragmentation. Within this large
category, Residential & commercial development and Agriculture &
aquaculture were widely studied using ML methods. Overall, most studies
on most threats used a common set of methodologies, except for
Pollution, for which the most used ML methods were ensembles, DTs and
ANNs.
3.2 Conservation measures
Similarly to threats, the most popular methods used to study
conservation measures were MaxEnt, Bayesian and ensembles (Fig. 2, Annex
E). These were mainly used in studies focusing on the effects of land
protection (Annex C), followed by species and land management.
The trends in the number of accumulated publications show the consistent
growing popularity of Bayesian methods, MaxEnt and ensemble methods
(Fig. 3). Other methods show a plateau in their popularity over the
years, which is most clear in evolutionary algorithms.