Resetting the baseline: using machine learning to find lost meadows

Abstract

Context: Mountain meadows occur in specific geomorphological conditions where low-gradient topography promotes fine sediment accumulation and high groundwater tables. Over 150 years of human-caused hydrological degradation of meadows along with fire suppression has resulted in decreased groundwater elevations and encroachment of upland vegetation, greatly diminishing the ecological value of meadows for water storage, baseflow, sediment capture, wildfire resistance, wildlife habitat, and carbon storage. Objectives: We aimed to understand where and how frequently meadows historically occurred to reset the baseline condition and provide insight into their restoration potential. We trained machine learning algorithms to identify potential meadow areas with similar hydrogeomorphic conditions to extant meadows while ignoring their unique vegetative characteristics because we hypothesized that vegetation would change but geomorphology would remain. Methods: We used a publicly available dataset of over 11,000 hand-digitized meadow polygons occurring within a 25,300 km2, 60-watershed region in the Sierra Nevada, California USA to train random forest models to detect meadow-like hydrogeomorphic conditions. Predictor variables represented topographical position, flow accumulation, snowpack, and topographical relief at differing spatial scales. We assessed model performance and produced maps delineating high probability meadow polygons. Results: Our findings showed that there is nearly three times more potential meadow habitat than currently documented. The predicted area includes a mixture of existing but undocumented meadows, non-meadowlands that may have converted from meadows due to lost function and forest encroachment, and areas with meadow-like geomorphology that may never have been meadow. The polygons encompassing predicted meadows often expanded existing meadows habitats into adjacent areas with continuous topography, but with upland vegetation and incised channels. Conclusions: Using readily available data and accessible statistical techniques, we demonstrated the accuracy of a tool to detect about three times more historical meadows than currently recognized within a complex, mountainous landscape. This “found” area greatly increased the potential area that could be subject to meadow restoration with benefits for biodiversity, wildfire management, carbon sequestration, and water storage.