SnowMicroPen Research Articles

Automatic snow type classification of snow micropenetrometer profiles with machine learning algorithms

Abstract. Snow-layer segmentation and classification are essential diagnostic tasks for various cryospheric applications. The SnowMicroPen (SMP) measures the snowpack's penetration force at submillimeter intervals in snow depth. The resulting depth–force profile can be parameterized for density and specific surface area. However, no information on traditional snow types is currently extracted automatically. The labeling of snow types is a time-intensive task that requires practice and becomes infeasible for large datasets. Previous work showed that automated segmentation and classification is, in theory, possible but cannot be applied to data straight from the field or needs additional time-costly information, such as from classified snow pits. We evaluate how well machine learning models can automatically segment and classify SMP profiles to address this gap. We trained 14 models, among them semi-supervised models and artificial neural networks (ANNs), on the MOSAiC SMP dataset, an extensive collection of snow profiles on Arctic sea ice. SMP profiles can be successfully segmented and classified into snow classes based solely on the SMP's signal. The model comparison provided in this study enables SMP users to choose a suitable model for their task and dataset. The findings presented will facilitate and accelerate snow type identification through SMP profiles. Anyone can access the tools and models needed to automate snow type identification via the software repository “snowdragon”. Overall, snowdragon creates a link between traditional snow classification and high-resolution force–depth profiles. Traditional snow profile observations can be compared to SMP profiles with such a tool.

Local-scale variability of snow density on Arctic sea ice

Abstract. Local-scale variations in snow density and layering on Arctic sea ice were characterized using a combination of traditional snow pit and SnowMicroPen (SMP) measurements. In total, 14 sites were evaluated within the Canadian Arctic Archipelago and Arctic Ocean on both first-year (FYI) and multi-year (MYI) sea ice. Sites contained multiple snow pits with coincident SMP profiles as well as unidirectional SMP transects. An existing SMP density model was recalibrated using manual density cutter measurements (n=186) to identify best-fit parameters for the observed conditions. Cross-validation of the revised SMP model showed errors comparable to the expected baseline for manual density measurements (RMSE = 34 kg m−3 or 10.9 %) and strong retrieval skill (R2=0.78). The density model was then applied to SMP transect measurements to characterize variations at spatial scales of up to 100 m. A supervised classification trained on snow pit stratigraphy allowed separation of the SMP density estimates by layer type. The resulting dataset contains 58 882 layer-classified estimates of snow density on sea ice representing 147 m of vertical variation and equivalent to more than 600 individual snow pits. An average bulk density of 310 kg m−3 was estimated with clear separation between FYI and MYI environments. Lower densities on MYI (277 kg m−3) corresponded with increased depth hoar composition (49.2 %), in strong contrast to composition of the thin FYI snowpack (19.8 %). Spatial auto-correlation analysis showed layered composition on FYI snowpack to persist over long distances while composition on MYI rapidly decorrelated at distances less than 16 m. Application of the SMP profiles to determine propagation bias in radar altimetry showed the potential errors of 0.5 cm when climatology is used over known snow density.

The RHOSSA campaign: multi-resolution monitoring of the seasonal evolution of the structure and mechanical stability of an alpine snowpack

Abstract. The necessity of characterizing snow through objective, physically motivated parameters has led to new model formulations and new measurement techniques. Consequently, essential structural parameters such as density and specific surface area (for basic characterization) or mechanical parameters such as the critical crack length (for avalanche stability characterization) gradually replace the semiempirical indices acquired from traditional stratigraphy. These advances come along with new demands and potentials for validation. To this end, we conducted the RHOSSA field campaign, in reference to density (ρ) and specific surface area (SSA), at the Weissfluhjoch research site in the Swiss Alps to provide a multi-instrument, multi-resolution dataset of density, SSA and critical crack length over the complete winter season of 2015–2016. In this paper, we present the design of the campaign and a basic analysis of the measurements alongside predictions from the model SNOWPACK. To bridge between traditional and new methods, the campaign comprises traditional profiles, density cutter, IceCube, SnowMicroPen (SMP), micro-computed-tomography, propagation saw tests and compression tests. To bridge between different temporal resolutions, the traditional weekly to biweekly (every 2 weeks, used in this sense throughout the paper) snow pits were complemented by daily SMP measurements. From the latter, we derived a recalibration of the statistical retrieval of density and SSA for SMP version 4 that yields an unprecedented spatiotemporal picture of the seasonal evolution of density and SSA in a snowpack. Finally, we provide an intercomparison of measured and modeled estimates of density and SSA for four characteristic layers over the entire season to demonstrate the potential of high-temporal-resolution monitoring for snowpack model validation.

Estimation of snow density from SnowMicroPen measurements

Snow density information is essentially required in snowpack research. We estimated snow density from penetration resistance measured with the SnowMicroPen (SMP). A 100cm3 density cutter was used for manual measurements of snow density. The measurements were taken at Patsio (3800m) station of SASE located in the Greater Himalayan range. The maximum penetration resistance values obtained for 1mm depth windows of SMP profiles were used for snow density modelling. A statistical density model is then proposed from these data which correlated well with the manual measurements (rp: 0.83). The model was tested for two other stations (Dhundi and Gulmarg) located in different climatic zones. A correlation of 82% for Dhundi and 83% for Gulmarg was found. The model was then compared with existing SMP based density models. The present model estimated densities fairly well for various snow types as compared to other SMP based models except for fresh snow. A new parameter, the number of peaks of maximum penetration resistance for 1mm depth (Np) has been introduced. The parameter is found to be uncorrelated with other statistical parameters presented so far in snow characterization.

Density, specific surface area, and correlation length of snow measured by high‐resolution penetrometry

AbstractPrecise measurements of snow structural parameters are crucial to understand the formation of snowpacks by deposition and metamorphism and to characterize the stratigraphy for many applications and remote sensing in particular. The area‐wide acquisition of structural parameters at high spatial resolution from state‐of‐the‐art methods is, however, still cumbersome, since the time required for a single profile is a serious practical limitation. As a remedy we have developed a statistical model to extract three major snow structural parameters: density, correlation length, and specific surface area (SSA) solely from the SnowMicroPen (SMP), a high‐resolution penetrometer, which allows a meter profile to be measured with millimeter resolution in less than 1 min. The model was calibrated by combining SMP data with 3‐D microstructural data from microcomputed tomography which was used to reconstruct full‐depth snow profiles from different snow climates (Alpine, Arctic, and Antarctic). Density, correlation length, and SSA were derived from the SMP with a mean relative error of 10.6%, 16.4%, and 23.1%, respectively. For validation, we compared the density and SSA derived from the SMP to traditional measurements and near‐infrared profiles. We demonstrate the potential of our method by the retrieval of a two‐dimensional stratigraphy at Kohnen Station, Antarctica, from a 46 m long SMP transect. The result clearly reveals past depositional and metamorphic events, and our findings show that the SMP can be used as an objective, high‐resolution tool to retrieve essential snow structural parameters efficiently in the field.

Comparison of micro-structural snowpack parameters derived from penetration resistance measurements with fracture character observations from compression tests

Compression tests are snow stability tests that are widely used by avalanche professionals and snow researchers to identify potential weak snowpack layers. Describing fracture character in addition to the number of taps required to initiate a fracture improves the interpretation of compression test results, since certain types of fractures, i.e. sudden fractures, are more often associated with skier-triggered avalanches. Digital snowpack penetrometers provide high resolution penetration resistance data of the snow cover with depth. The SnowMicroPen (SMP) was used to measure high resolution penetration resistance profiles in the snowpack next to compression tests. A reliable method to automatically detect the snow surface in the SMP signals was introduced. Furthermore, a method based on the autocorrelation of the penetration resistance signal was developed to identify the failure layers, identified using compression tests, in the penetration resistance profiles. Using field data from 190 penetration resistance measurements, each collected between two compression tests, micro-structural parameters associated with different types of fractures were identified. More than 550 fractures were classified as either Progressive Compression (1.3%), Resistant Planar (12.1%), Sudden Planar (50.4%), Sudden Collapse (26.8%) or non-planar Break (9.4%). Measurement and analysis were focussed on micro-structural properties of the failure layer, the layer adjacent to the failure layer and the slab above the failure layer. Sudden collapse fractures were found to have typical micro-structural snowpack parameters that are generally associated with unstable snow conditions, such as large differences in penetration resistance between the failure layer and the adjacent layer.

Rutschblock-scale snowpack stability derived from multiple quality-controlled SnowMicroPen measurements

Stability prediction from SnowMicroPen (SMP) profiles would support avalanche forecasting operations, since objective stability information could be gathered more quickly than with standard tests, thereby allowing sampling at higher resolution and over larger spatial scales. Previous studies have related the snow properties derived from the SMP to observed snow properties at Rutschblock (RB) and compression test failure planes. The goals of this study are to show to what extent snowpack stability for artificial triggering, based on RB, can be derived from SMP measurements and how multiple measurements at the RB scale improve the results. Measurements at 36 different sites were used for the development of a classification scheme. Each site included a RB test, a manual profile, and 6 to 10 adjacent SMP measurements, for a total of 262 SMP profiles. A recently improved SMP theory was applied to estimate the micro-structural and mechanical properties of manually defined weak layers and slab layers. SMP signal quality control and different noise treatment methods were taken into consideration in the analysis. The best and most robust predictor of RB stability was the weak layer micro-scale strength. In combination with the SMP-estimated mean density of the slab layer, the total accuracy of predicting RB stability classes was 85% over the entire dataset, and 88% when signals with obvious signal dampening (11% of the dataset) were removed. The total accuracy increased when multiple SMP measurements at the RB scale were used to calculate the mean weak layer strength, when compared to using just one SMP measurement at a site. The analysis was robust to trends and offsets in the absolute SMP force, which was a frequent signal error. However, it was sensitive to dampened or disturbed SMP force micro variance. The sensitivity analysis also showed that the best predictor of instability, the weak layer micro-scale strength, was robust to the choice of SMP signal noise removal method.

Quantifying changes in weak layer microstructure associated with artificial load changes

Researchers and practitioners have long utilized a variety of penetrometers to investigate the snowpack. Identifying definitive relationships between penetrometer-derived microstructural information and stability has been challenging. The purpose of this study is two-fold: 1. We propose a simple field test to establish relationships between load and penetrometer-derived microstructural estimates, 2. We utilize the SnowMicroPen (SMP) to quantify changes in weak layer residual strength and microstructural dimension associated with an artificial loading event. Our dataset is from Moonlight Basin, Montana and includes three modified loaded-column tests, each paired with 5 SMP profiles. Depth hoar comprised the targeted weak layer. Results indicate that loading caused the residual strength and rupture frequency to decrease significantly. Much like a compression test at a micro-scale, the force required for the SMP to rupture individual structures as well as the micro-scale strength decreased significantly when the slab stress was increased by artificially adding blocks of snow. A decrease in observed rupture frequency within the weak layer (or an increase in the distance between ruptured structures) also occurred after the loading event, probably because some structures within the weak layer had already failed or were so close to failing that the penetrometer could not detect their rupture. Due in part to the large difference in loads, microstructural differences between the natural and loaded columns were significant enough that only one profile would have been necessary to determine a significant difference in residual strength. Artificial removal of slab stress resulted in greater rupture forces and larger microstructures, likely due to elastic rebound.

Stability algorithm for snow micro-penetrometer measurements

AbstractInformation on snow-cover stability is important for predicting avalanche danger. Traditionally, stability evaluation is based on manual observations of snow stratigraphy and stability tests, which are time-consuming. The SnowMicroPen (SMP) is a high-resolution, constant-speed penetrometer to measure penetration resistance. We have analysed the resistance signal to derive snow stability. The proposed stability algorithm was developed by comparing 68 SMP force–distance profiles with the corresponding manual profiles, including stability tests. The algorithm identifies a set of four potentially weak layers by taking into account changes in structure and rupture strength of microstructural elements that make up snow layers as derived from the SMP signal. In 90% of the cases, one of the four potentially weak layers proposed by the algorithm coincided with the failure layer observed in the stability test. To select the critical layer from the four potential weaknesses was more difficult. With fully automatic picking of the critical layer, agreement with the failure layer observed in the stability test was reached in 60% of the cases. To derive a stability classification, we analysed weak-layer as well as slab properties. These predictor variables allow the SMP signal to be classified into two stability classes, poor and fair-to-good, with an accuracy of ∼75% when compared with observed stability. The SMP, in combination with the proposed algorithm, shows high potential for providing snow-cover stability information at high resolution in time and space.

Impact of the microstructure of snow on its temperature: A model validation with measurements from Summit, Greenland

The influence of snow microstructure on thermal and radiative transfer in snow has not been thoroughly investigated as the tools necessary to efficiently measure microstructural geometry at millimeter resolution have not yet been available. Here we investigate the impact of snow microstructure on the thermal and radiative properties of snow and specifically determine the depth resolution of snow stratigraphy measurements needed to adequately model snow temperatures. To address this subject, detailed information on physical properties of the snow cover was collected at Summit, Greenland, in summer 2003. We present a new set of snow microstructure data, measured with a high‐resolution penetrometer (SnowMicroPen (SMP)). The penetration resistance can be used to estimate the thickness of individual layers but also reflects the thickness and the number of bonds in the snowpack. SMP is a motor‐driven, constant speed micropenetrometer acquiring 256 hardness measurements per millimeter. The data were used to describe the physical properties of snow and to compute snow temperatures in the topmost meter of the snowpack using the snow cover model SNOWPACK. The spatial resolution of the snow microstructure and the subsequent parametrization of SMP‐estimated density strongly affected the modeled temperature and temperature gradients. The SMP‐estimated density was used as an input to SNOWPACK and affected the thermal conductivity and radiative transfer in the snowpack. Our results thus show that highly resolved stratigraphic input data on the order of at least 1 cm are necessary to adequately simulate snow temperatures and snow metamorphism, including the observed formation of subsurface hoar in the high Arctic snow cover.

Surface hoar characteristics derived from a snow micropenetrometer using moving window statistical operations

This study aims to improve analytical techniques for studying stratigraphic dimensions and hardness characteristics of thin weak layers in the mountain snowpack, with particular interest to buried surface hoar layers. By determining which structural characteristics of such weak layers are associated with shear strength, we may be better able to monitor and predict stability, which is relevant for avalanche forecasting and management. We utilize moving window statistical operations to analyze SnowMicroPen (SMP) penetrometer hardness profiles of a buried surface hoar layer. Results indicate that significant weak layer thinning and hardening of the interface between the weak layer and its substratum coincided with significant increases in shear strength, as measured using a size-corrected shear strength index derived from concurrent stability tests. With aging, variations in slab thickness appeared to positively affect the hardness and inversely affect the coefficient of variation (CoV) of hardness of weak layer boundaries. These findings support previous research that proposed the strengthening of buried surface hoar layers results from the gradual penetration of the surface hoar crystals into the substratum which allows stronger bonds to form at this critical interface. These analytical techniques allow stratigraphic dimensions and hardness characteristics to be quantified and analyzed, improving our ability to monitor stratigraphic characteristics associated with shear strength and stability of the mountain snowpack.

Snowpack stability information derived from the SnowMicroPen signal

Snowpack measurements and stability tests are, next to recent avalanche activity and weather history, currently the basis for snowpack stability assessment in most avalanche warning operations. The SnowMicroPen (SMP), a high-resolution penetrometer for snow, measures penetration resistance force or snow hardness. In order to be useful for an avalanche warning service, stability information needs to be provided and must be derivable from the SMP signal. SMP profiles (25 on slopes, 14 on flat sites) were taken together with manual snow profiles and stability tests, such as Rutschblock and compression tests. The data are from three winter seasons of the years 2001–2002 to 2003–2004 in the Swiss Alps. According to their stability test score and failure interface properties the manual profiles were classified as stable or unstable. Based on the manual observations the failure interfaces were identified in the SMP profiles and possible indicators of instability were derived from the SMP signals at these interfaces. The distinct indicators of instability were the failure layer micro-structural length and hardness, the difference in structural length across the failure interface and the failure layer macro-elastic modulus. The cross-validated accuracy of classification into stable or unstable failure interfaces gained from SMP parameters was comparable to the classification accuracy from manual profile parameters (about 65%). It remains to be tested if stability information can be derived from a SMP measurement without knowing the location of the failure interface found by a stability test. If this can be done successfully and reliably, avalanche warning operations could definitely benefit from the instrument.

Changes in the shear strength and micro-penetration hardness of a buried surface-hoar layer

AbstractWe investigated a buried surface-hoar layer using the SnowMicroPen (SMP), an instrument designed to measure detailed snowpack profiles.We collected data from two adjacent parts of a slope 6 days apart. In addition, one manual snowpack profile was sampled each day, as well as 50 quantified loaded column tests (QLCTs) which provided an index of shear strength. For the SMP data, a 900 m2 area was sampled on both days in a grid with points 3 mapart, with some sub-areas of more closely spaced measurements. We collected 86 SMP profiles on the first day and 129 on the second day. Our analyses involved manually locating layer boundaries and calculating statistics for the force signal through the surface-hoar layer. The shear strength index increased by 40% between the two sampling days, but the SMP data show no statistical difference in layer thickness, and the mean, minimum, median, and a variety of percentile measures of the SMP force signal through the layer also do not change. Interestingly, the maximum hardness, and the variance and coefficient of variation of the SMP signal, increased. Since the small SMP tip might only break one or a couple of bonds as it passes through the weak layer, we interpret these changes as being indicative of increasing bond strength. Though we cannot specifically tie the increasing maximum hardness of the SMP signal to our QLCT results, our work suggests that the maximum SMP signal within buried surface-hoar layers may be useful for tracking increases in the shear strength of those layers.