Virtual snow stakes: a new method for snow depth measurement at remote camera stations

Abstract

Remote cameras are used to study demographics, ecological processes, and behavior of wildlife populations. Cameras have also been used to measure snow depth with physical snow stakes. However, concerns that physical instruments at camera sites may influence animal behavior limit installation of instruments to facilitate collecting such data. Given that snow depth data are inherently contained within images, potential insights that could be made using these data are lost. To facilitate camera‐based snow depth observations without additional equipment installation, we developed a method implemented in an R package called edger to superimpose virtual measurement devices onto images. The virtual snow stakes can be used to derive snow depth measurements. We validated the method for snow depth estimation using camera data from Latah County, Idaho, USA in winter 2020–2021. Mean bias error between the virtual snow stake and a physical snow stake was 5.8 cm; the mean absolute bias error was 8.8 cm. The mean Nash Sutcliffe Efficiency score comparing the fit of the 2 sets of measurements within each camera was 0.748, indicating good agreement. The edger package provides researchers with a means to take critical measurements for ecological studies without the use of physical objects that could alter animal behavior, and snow data at finer scales can complement other snow data sources that have coarser spatial and temporal resolution.

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

“…Our model had measurement errors that were comparable to those of other methods used to measure snow depth (lidar ∼± 10 cm, manual snow probing ∼1-2 cm) ( Holmgren, 2018). Previous automated and semi-automated methods for obtaining snow depth from time-lapse imagery report accuracy within 4 and 10 cm (Bongio et al, 2021;Garvelmann et al, 2013;Strickfaden et al, 2023), and our model trained on the Colorado data set shows that this keypoint detection model performs with high accuracy and across more poles when presented with the pole site during training. The performance corresponds to the success of other keypoint models in detecting face features and estimating poses (K. Wang et al, 2010).…”

“…Striped poles are also advantageous because they provide intermediate depth markings to compare to the depth in pixels (Garvelmann et al, 2013;Hofmeester et al, 2019). Future iterations of the model could even incorporate virtual snow pole approaches, such as in Strickfaden et al, 2023, and superimpose the virtual snow pole onto the image of interest and select the top and bottom of the virtual pole for the training data. We call for future work to integrate multiple data sets from various ecosystems, multiple years, and pole set-ups into one model, thereby leveraging a larger sample size to improve accuracy on new data sets and increasing applicability to currently existing data sets.…”

“…Parajka et al (2012) presented a "routine" to automate the process, although the method was not described in detail. Strickfaden et al (2023) developed code to extract depth from virtual snow poles using reference objects, but snow depths must be read manually. Thus, these studies suggest tools for automation, but no study provides a scalable method at high precision.…”