Local land-use decisions drive losses in river biological integrity to 2099: Using machine learning to disentangle interacting drivers of ecological change in policy forecasts

Abstract

Climate and land-use/land-cover (LULC) change each threaten the health of rivers. Rising temperatures, changes in rainfall and runoff, and other perturbations, will all impact rivers' physical, biological, and chemical characteristics over the next century. While scientists and policymakers have increasing access to climate and LULC forecasts, the implications of each for outcomes of interest have been difficult to quantify. This is partially because climate and LULC perturb ecological outcomes via incompletely understood site-specific, interacting, and nonlinear mechanisms that are not well suited to analysis using classical statistical methods. This creates uncertainties over the benefits of local-level interventions such as green infrastructure investments and urban densification, and limits how forecasts can be used to inform decision-making. Here, we demonstrate how machine learning can be used to quantify the relative contributions of LULC and climate drivers to impacts on riverine health as measured by taxonomic richness of the macroinvertebrate orders Ephemeroptera, Plecoptera, and Trichoptera (EPT). We develop a cross-validated Random Forest (RF) model to link EPT taxa richness to meteorological, water quality, hydrologic, and LULC variables in watersheds in New Hampshire and Vermont, USA. Prospective climate and LULC scenarios are used to generate predictions of these variables and of EPT taxa richness trends through the year 2099. The model structure is mechanistically interpretable and performs well on test data (R² ~ 0.4). Impacts on EPT taxa richness are driven by local LULC policy such as increased suburbanization. Future trends are likely to be exacerbated by climate change, although warming conditions suggest possible increases in springtime EPT taxa richness. Overall, this analysis highlights (1) the impact of local LULC decisions on riverine health in the context of a changing climate, and (2) the role machine learning methods can play in developing models that disentangle interacting physical mechanisms to advance decision support.