Automatic Grain Type Classification of Snow Micro Penetrometer Signals With Random Forests

Snow microstructure plays an important role in the remote sensing of snow water equivalent (SWE) for both passive and active microwave radars. The accuracy of microwave SWE retrieval algorithms is sensitive to (usually unknown) changes in microstructure. These algorithms could be improved with high-resolution estimates of microstructural properties by using an advanced instrument such as the Snow Micro Penetrometer (SMP), which measures penetration force at the millimeter scale and is sensitive to microstructure. The SMP can also take full micromechanical measurements at much greater speed and resolution and without observer bias than a traditional snow pit. Previous studies have shown that the snowpack stratigraphy and grain type can be accurately classified with one SMP measurement using basic statistics and classification trees (CTs). For this study, we used basic statistical measures of the penetration force and micromechanical estimates from an SMP inversion algorithm to significantly improve the classification accuracy of grain type and layer discrimination. We applied random forest (RF) techniques to classify three snow grain types (new snow, rounds, and facets) from SMP measurements collected in Switzerland and Grand Mesa, Colorado. RFs performed up to 8% better than single CTs, with overall misclassification errors between 17% and 40%. The coefficient of variation of the penetration force proved to be the most important variable, followed by variables that contain information about grain size like microscale strength and the number of ruptures.