Mon, Aug 15, 2022: 4:30 PM-4:45 PM
516B
Background/Question/MethodsSelecting an appropriate plant species mix for a site is fundamental to the success of a restoration effort. When available, practitioners may draw inspiration from intact reference sites with environmental conditions resembling the target. Yet, in systems where anthropogenic disturbance is widespread, such as agriculturally valuable semi-arid grasslands, few if any intact references may remain. Other avenues include consulting herbaria records for a region of interest, though these are rarer and spatially imprecise for those observations that precede industrialized agriculture. We explored alternative methods of species selection by gathering repeat vegetation survey data, records of agricultural disturbance, and gridded environmental data for sites in the Bitterroot Valley, MT, USA. We then trained a machine learning model to predict species-level forb abundance. A key objective was to estimate the negative influences that cultivation and cattle ranching had on the plant community. This is a challenge because predictive models may misattribute the negative effects of agriculture to environmental properties that correlate with anthropogenic effects. To address this, we employed Shapley Additive Explanations (SHAP; https://github.com/slundberg/shap) to reveal whether our model attributed positive or negative influences to these variables.
Results/ConclusionsAcross 99 native forb species we found that a model naive to agricultural disturbance assigned negative effects to low elevation and flatter sites, as compared to a disturbance-informed model where the effects of these conditions were more neutral and agricultural effects strongly negative. These findings suggest that a model naive to agricultural disturbance may learn proxies for these effects from interactions among elevation and slope in our system, while a disturbance-informed model properly attributes these effects to disturbance variables. Performance across the two models was similar, yet a disturbance informed model offers a key advantage in that it enables spatially explicit simulations of plant abundance where agricultural effects are nullified and other conditions held constant. We discuss how this information may be cautiously employed to guide plant materials trials and aid development of species mixes in areas lacking intact reference conditions.
Results/ConclusionsAcross 99 native forb species we found that a model naive to agricultural disturbance assigned negative effects to low elevation and flatter sites, as compared to a disturbance-informed model where the effects of these conditions were more neutral and agricultural effects strongly negative. These findings suggest that a model naive to agricultural disturbance may learn proxies for these effects from interactions among elevation and slope in our system, while a disturbance-informed model properly attributes these effects to disturbance variables. Performance across the two models was similar, yet a disturbance informed model offers a key advantage in that it enables spatially explicit simulations of plant abundance where agricultural effects are nullified and other conditions held constant. We discuss how this information may be cautiously employed to guide plant materials trials and aid development of species mixes in areas lacking intact reference conditions.