Vegetation controls on sub-canopy snow depth variation

Abstract:

Vegetation is an influential driver of sub-canopy snow depth with direct (interception from vegetation structure) and indirect (wind and shortwave radiation shelter) effects. Vegetation predictors are not necessarily linearly related to sub-canopy snow depth and advanced statistical methods may address the complex relationships between them. This study aims to measure the effect of vegetation structure on sub-canopy snow depths through machine learning models. In this regard, we use four supervised machine learning approaches (Support Vector Machine, Random Forest, Neural Network and Elastic Net) to estimate sub-canopy snow depth from principal component analysis (PCA) transformed vegetation metrics in a nonlinear regression problem. Vegetation metrics are derived from individual trees of 6 terrestrial laser scanning (TLS) sites (A, F, K, M, N and O) collected across Grand Mesa, Colorado during the 2016-2017 and 2019-2020 NASA SnowEx campaigns. Here we present the results from the 2016-2017 dataset, and will provide the results from the 2019-2020 dataset at the time of presentation. As snow processes can be site specific, the models are applied for each site separately. To avoid overfitting we use bootstrap sampling, and model parameters are estimated using a grid search on a range of hyperparameters. Our results from the 2016-2017 dataset indicate that the first two PCA components cover more than 70% of the variance in vegetation metrics, which results in more than 80% accuracy in estimating snow depths in the sub-canopy. We apply the four models on sites altogether to examine the total vegetation effect on sub-canopy snow depth variation. The results show overfitting in the RF model but the other three models predict up to 80%of sub-canopy snow depths. Site specific models also represent up to 90% accuracy on a test dataset. In addition, among the four models, Neural Network and Support Vector Machine show the highest accuracy for both individual and overall sites.