Search ETDs:
Estimating snow depth of alpine snowpack via airborne multifrequency passive microwave radiance observations
Kim, Rhae Sung

2017, Doctor of Philosophy, Ohio State University, Geodetic Science.
Snow cover plays a key role in the climate and water resource systems in mountainous areas; therefore, accurately monitoring snow properties (e.g., snow water equivalent (SWE) or snow depth) is critical. Although snow depth can be estimated in-situ, these measurements are expensive and generally limited in spatial coverage. Other methods, namely snow hydrologic modeling and remote sensing, have their intrinsic strengths and limitations; accurate knowledge and understanding of their highly complementary relations are required.
In this study, we utilized passive microwave (PM) measurements of the brightness temperature (Tb) to characterize snowpack properties in mountainous areas. Tb exhibits reduced sensitivity to depth for deep snow and in forests, limiting the ability of many existing algorithms for snow mapping. An alternative approach is to classify snow depth based on its multifrequency Tb signatures. Here, we first analyzed airborne Tb measurements of alpine snowpack for five frequencies and two polarizations, and compared them with an estimate of forest cover and concurrent measurements of snow depth and snow wetness collected as part of the NASA Cold Land Processes Field Experiment. We analyzed a total of 900 independent samples, each representing one hectare. Samples were classified into classes based on snow depth, forest fraction, and wetness. We assessed whether the mean Tb spectrum of each class differed from other classes using the Hotelling's T-squared test, and assessed the separability of classes using the Jefferies-Matusita (J-M) distance. Hotelling's T-squared test revealed that the Tb for each forest cover and snow depth class differed statistically from each of the others, for dry snow, notwithstanding that within-class Tb variability tended to be larger than the between-class differences. The J-M distance indicated that most classes were somewhat separable based on the Tb spectra. Consistent with expectations, J-M distance between classes was lower for forested areas than for un-forested areas, emphasizing the confounding influence of trees on characterizing snow using Tb measurements. Based on the results of separability tests, we explored the supervised machine learning approach by using various classifiers and RBF-SVM (Support Vector Machine with RBF kernel function) was selected with highest accuracy. In our classification system, we utilized both vertical and horizontal polarizations of Tb in order to provide maximal information to the classification predictor. Classification accuracy was compared with the accuracy when using only Tb at vertical polarization. Classification accuracies tended to decrease with increasing forest cover density; however, it was encouraging that snow depth could be somewhat classified even when pixels were forested. Classification results for all different forest cover conditions showed improved overall accuracies when using both horizontal and vertical polarizations instead of using only vertical polarization.
Based on a study of Tb spectra, we proposed a new snow depth retrieval algorithm for mountainous deep snow using airborne multifrequency PM radiance observation. In contrast to previous snow depth estimations using satellite PM radiance assimilation, the newly- proposed method utilized a single flight observation and deployed the snow hydrologic models as a basis for a “snapshot” retrieval algorithm. This method is promising since the satellite-based retrieval methods have difficulties to estimate snow depth due to their coarse resolution and computational effort. Our approach consists of a particle filter using combinations of multiple PM frequencies and multi-layer snow physical model (i.e., Crocus) to resolve melt-refreeze crusts. Results showed that there was a significant improvement over the prior snow depth estimates and the capability to reduce the prior snow depth biases. When applying our snow depth retrieval algorithm using a combination of four PM frequencies (10.7-, 18.7-, 37.0-, and 89.0 GHz), the root mean square error (RMSE) values were reduced by 62% at the snow depth transects sites where forest density was less than 5% despite deep snow conditions. This method displayed a higher sensitivity to different combinations of frequencies, model stratigraphy (i.e. different number of layering scheme for snow physical model) and estimation methods (particle filter and Kalman filter) except the forest cover density and precipitation bias. The prior RMSE values at the forest-covered areas were reduced by 27 - 41% even in the presence of forest cover.
Michael Durand (Advisor)
Alper Yilmaz (Committee Co-Chair)
Ralph Von Frese (Committee Member)
136 p.

Recommended Citations

Hide/Show APA Citation

Kim, R. (2017). Estimating snow depth of alpine snowpack via airborne multifrequency passive microwave radiance observations. (Electronic Thesis or Dissertation). Retrieved from https://etd.ohiolink.edu/

Hide/Show MLA Citation

Kim, Rhae Sung. "Estimating snow depth of alpine snowpack via airborne multifrequency passive microwave radiance observations." Electronic Thesis or Dissertation. Ohio State University, 2017. OhioLINK Electronic Theses and Dissertations Center. 16 Oct 2018.

Hide/Show Chicago Citation

Kim, Rhae Sung "Estimating snow depth of alpine snowpack via airborne multifrequency passive microwave radiance observations." Electronic Thesis or Dissertation. Ohio State University, 2017. https://etd.ohiolink.edu/

Files

PhD.Dissertation_RSK .pdf (2.69 MB) View|Download