Machine learning to reconstruct atmospheric river landfalls along the US West Coast using tree-ring based moisture proxies

Atmospheric Rivers (ARs) are a major component of hydrometeorology in Western North America. They contribute substantially to annual precipitation totals but also drive extreme precipitation and flooding events. The central role of ARs has spurred significant interest in how these events might change under anthropogenic warming, which in turn requires an understanding of their natural variability to help contextualize observed and projected trends. This study focuses on characterizing the natural variability of the frequency and landfall location of ARs by developing a reconstruction of AR landfalls across the US West Coast based on tree-ring based moisture proxies from across the Western US. These reconstructions leverage the strong relationship between AR landfalls and cold-seasonal precipitation recorded by tree-ring chronologies, both along the coast and further into the Inter-Mountain West. We utilize a gridded, tree-ring based reconstruction of cold-season precipitation across the Western US to develop a multi-output neural network (MNN) that predicts the annual counts of AR landfalls at six locations along the coastline. The MNN is designed to account for varying spatial patterns of tree-ring recorded precipitation associated with different AR pathways into the continental interior. We use NN dropout to regularize the model and construct uncertainty intervals, and employ saliency maps to identify the locations of tree-ring based moisture proxies that best support the reconstruction. These approaches help to reduce the "black-box" nature of the MNN and support a critical appraisal of the relationships identified in the model. The reconstruction is extended back to 1400 CE to help better understand the latitudinal variability of ARs over the last millennium, including low-frequency variability that could obscure trend detection in the latitude of AR landfalls.