Machine learning and hurdle models for improving regional predictions of stream water acid neutralizing capacity

Abstract

[1] In many industrialized regions of the world, atmospherically deposited sulfur derived from industrial, nonpoint air pollution sources reduces stream water quality and results in acidic conditions that threaten aquatic resources. Accurate maps of predicted stream water acidity are an essential aid to managers who must identify acid-sensitive streams, potentially affected biota, and create resource protection strategies. In this study, we developed correlative models to predict the acid neutralizing capacity (ANC) of streams across the southern Appalachian Mountain region, USA. Models were developed using stream water chemistry data from 933 sampled locations and continuous maps of pertinent environmental and climatic predictors. Environmental predictors were averaged across the upslope contributing area for each sampled stream location and submitted to both statistical and machine-learning regression models. Predictor variables represented key aspects of the contributing geology, soils, climate, topography, and acidic deposition. To reduce model error rates, we employed hurdle modeling to screen out well-buffered sites and predict continuous ANC for the remainder of the stream network. Models predicted acid-sensitive streams in forested watersheds with small contributing areas, siliceous lithologies, cool and moist environments, low clay content soils, and moderate or higher dry sulfur deposition. Our results confirmed findings from other studies and further identified several influential climatic variables and variable interactions. Model predictions indicated that one quarter of the total stream network was sensitive to additional sulfur inputs (i.e., ANC < 100 µeq L−1), while <10% displayed much lower ANC (<50 µeq L−1). These methods may be readily adapted in other regions to assess stream water quality and potential biotic sensitivity to acidic inputs.