Unpacking the "black box": improving ecological interpretation of regression based models

AimMany tree species distribution models use black-box machine learning
techniques that often neglect interpretative aspects and instead focus
mainly on maximising predictive accuracy. In this study, we outline an
interpretative modelling framework to gain better ecological insights
while mapping abundance patterns of six North American species.
LocationContinental United States and Canada MethodsWe develop an
innovative procedure using regression trees by stabilising variance and
mapping dominant rules which we term ‘optimized regression tree bagging
for interpretation and mapping’ (ORTBIM). We apply this technique to
understand ecological features influencing the abundance patterns of three
eastern (Pinus strobus, Acer saccharum, and Quercus montana), and three
western (Picea engelmannii, Pinus ponderosa, and Pseudotsuga menziesii)
tree species in North America. For these species, we assess and map the
dominant climate-terrain interactions that partly determine abundance
patterns in the eastern and western regions. In the process, we examine
the role of varying responses and scales and explore finer-scale species
climate-terrain niches and non-linear relationships. ResultsOur study
emphasizes the prominent role of elevation and heat-moisture variables in
the west and the greater importance of seasonal precipitation and seasonal
temperature in the east. The abundance patterns under future climate
(SSP5–8.5) show climate-terrain habitats shifting northward and westward
into Canada and Alaska for the eastern species, and predominantly
north-westward for the western species. ConclusionOur interpretative
modelling framework can be used to gain a more comprehensive understanding
of the abundance patterns across the full species range, to formulate
better predictive models, and to facilitate improved management practices
under climate change.