Snow Depth Extraction From Time‐Lapse Imagery Using a Keypoint Deep Learning Model

Abstract

Snow pole time‐lapse photography, in which a snow pole of a known height is installed in front of a camera and photographed repeatedly over the course of a snow season, allows a large network of sites to be established relatively quickly and affordably. However, current approaches for extracting snow depth from snow poles typically relies on time intensive manual photo processing. By integrating computer vision algorithms with snow pole photography, we present a method that uses a keypoint detection model to automatically observe snow height across a network of sites. At 20 snow pole locations from Grand Mesa, CO (n = 9,722 images), our model successfully predicts the top and bottom of the pole with a mean absolute error (MAE) of 1.30 cm. To assess model generalizability, we tested the model on 12 sites in Washington State (n = 1,770 images). When the Colorado trained model was fine‐tuned using a subset of Washington images, the model predicted snow depth with a MAE of 4.0 cm. Best performance was achieved when both data sets were included during training, with a MAE of 2.05 cm for Colorado images and a MAE of 1.14 cm for Washington images. We demonstrate that, especially when trained using a subset of site‐specific data, a keypoint detection model can accelerate snow pole automation. This algorithm brings the hydrology community one step closer to a generalized snow pole detection model, and we call for a future model that integrates across time‐lapse images from additional locations.