Computer Vision tools for image feature characterization, recognition and predictability in high-resolution snow depth distributions

The efficacy of physically based models deployed in snow-affected environments relies heavily on the accuracy of spatially distributed precipitation inputs. The use of spatially distributed snow depth data to help constrain these precipitation inputs for improved model forcing data is very appealing, particularly when highly accurate products such as those from ASO (Airborne Snow Observatory) are available. While it has been shown that snow depth patterns can exhibit spatial persistence in a retrospective setting, which matches what our brains intuitively see, we do not have a clear path to how this information might be captured and employed in forward-looking applications.

We applied the concepts of pattern recognition and feature extraction, as developed in the discipline of computer vision, to identify and characterize spatial features in snow depth scenes from ASO data across multiple locations. These spatial features were then used to compare snow depth patterns across intra-annual and inter-annual timescales to assess the similarity, persistence and predictability of snow depth patterns. With the use of computer vision tools, we can circumvent spatial-autocorrelation concerns that are associated with more traditional correlation-based statistics to quantify the similarity in spatial distribution. In addition, the automated extraction of features from snowpack distributions mimic those assessments that are quickly identified by the human eye. The extracted features provide unique feature-based insight into the study of snow persistence and spatial snowpack distribution.

In this study, we highlight that both spatial pattern and magnitude of the spatial variability within the snow scene are required to constrain spatially distributed precipitation data. Furthermore, these amplitudes and spatial gradients may be subject to considerable inter-annual variability, thus leaving significant challenges for forward-looking predictions in operational model settings, even in areas with many ASO collections.