Flipping through field notebooks…
Abstract Accurately estimating snow water equivalent (SWE) is crucial for understanding the impacts of climate change, urbanization, and population growth on water resources. High operational costs of lidar observations limit the frequency and coverage of SWE estimates at high spatial resolutions, leading to significant data gaps. We address this challenge with a transfer learning (TL) framework that leverages abundant SWE data from California to enhance predictions in Colorado, where data are scarce. From 2016 to 2019, the disparity in SWE data collection between these states was stark: 94 snowpack maps were recorded in California's Sierra Nevada versus only 12 in Colorado's Rocky Mountains. We hypothesized that geographic predictors (e.g., elevation and snowfall) would exhibit similar effects on SWE across these landscapes. By conducting an exploratory factor analysis, we validated this hypothesis and refined our TL model, which incorporated data based on 80 snowpack maps from California to predict SWE in Colorado. The analysis included six models: Local 1, Local 1W, TL1, TL2, TL1W, and TL2W. Local models were trained only on Colorado data, while TL models leveraged TL using a base model trained on California data. Weighted models (Local 1W, TL1W, TL2W) incorporated a weighting mechanism to emphasize critical predictors. Compared to using data from Colorado alone, TL improved the mean value from 0.45 to 0.54, representing a significant enhancement of 20% in predictive accuracy, while reducing bias from 0.24 to 0.17 (a 30% decrease). Our framework helps mitigate lidar data limitations, supporting water management.
Knowledge graph centered on Modeling spatial distribution of snow water equiva with 4 nodes and 4 connections. Top connected: Dipankar Dwivedi, Utkarsh Mital, L. E. Halabi.
Items connected by shared entities, co-authorship, citations, or semantic similarity.
51 references to works outside the Knowledge Commons