Evaluating L-band radar for the future of snow remote sensing

Abstract

Snowpack monitoring is essential because seasonal snowpacks provide water for billions of people, support streamflow and ecosystems, and are a fundamental component of the Earth's energy system. However, no current snowpack monitoring system is capable of measuring snow water equivalent (SWE), the most important snowpack hydrologic variable, accurately and at high spatiotemporal (<500 m, <weekly) resolutions. The primary goal of this dissertation is to evaluate a method for deriving snow density using a ground-based radar and a method for retrieving SWE from an airborne platform. In Chapter 2, I study the spatial variability of snow density by deriving relative permittivity, and thereby bulk density, from combined ground-penetrating radar (GPR) and lidar measurements of the snowpack. In Chapters 3–4, I evaluate the L-band InSAR technique for SWE retrievals, which has been considered to be a promising method for SWE retrievals for more than a decade, but has seen limited testing. I evaluate the technique using repeat InSAR acquisitions from the airborne NASA UAVSAR instrument in meadow and burned environments at Cameron Pass, Colorado (Chapter 3) and in the montane forest environment of Fraser Experimental Forest, Colorado (Chapter 4). Snow density is a critical input variable for the retrieval of SWE from most SWE remote sensing methods. However, density is time consuming to measure in the field and is thus often measured sparsely, preventing extensive analysis of snow density models, a primary source of estimated densities for remote sensing methods. In Chapter 2, I derived >20 km of nearly continuous relative permittivity estimates, and thereby bulk density, from combined near-coincident measurements of GPR two-way travel times and lidar snow depths at three different field sites and in both dry and wet snow conditions. Variogram analyses were conducted and revealed a 19 m median correlation length for relative permittivity and density in dry snow. For wet snow, the correlation length increased to >30 m. I then leveraged the derived densities to evaluate six snow density models to better understand the limitations of these models within lidar and radar remote sensing methods. Two models yielded densities that estimated SWE within ±10% when SWE exceeded 400 mm, but model uncertainty increased to >20% when SWE was less than 300 mm. Thus, the refinement of these density models and the development of future density models is a high priority to fully realize the potential of SWE remote sensing methods. The L-band (1–2 GHz) InSAR technique for measuring changes in SWE (ΔSWE) is a promising method for SWE retrievals because the longer wavelength (~0.25 m) has minimal interaction with the snowpack microstructure and has increased canopy penetrative capabilities. In Chapter 3, I evaluated 10 L-band InSAR pairs collected by NASA UAVSAR near Cameron Pass, Colorado with GPR and terrestrial lidar measurements of ΔSWE in open meadows and burned forests. For single InSAR pairs, UAVSAR ΔSWE retrievals yielded an overall Pearson's correlation coefficient of 0.72–0.79, with a RMSE of 19–22 mm. I expanded the analysis beyond the locations of GPR and lidar surveys to evaluate the time series of UAVSAR SWE retrievals by including measurements of SWE from seven automated stations and found a RMSE of 42 mm. These findings support the use of this technique in unforested areas with dry snow conditions for the upcoming L-band NISAR satellite mission Given the findings of Chapter 3 and the canopy penetration capabilities of L-band radar, I designed Chapter 4 to evaluate the influence of forest cover on the UAVSAR signal. In Chapter 4, I evaluated eight L-band InSAR pairs collected by UAVSAR over the montane forests of Fraser Experimental Forest, Colorado with manually surveyed snow depths and snow pits and a pair of airborne lidar surveys. Compared with in situ measurements, I found that forest cover fractions <40% yielded RMSEs of ~15 mm, whereas RMSE more than doubled for forest cover fractions >50%. Further, normalized cumulative UAVSAR SWE and normalized lidar snow depths yielded identical statistical distributions for forest cover fractions <50% across the full study area, but these distributions diverged as forest cover fraction increased. Thus, forest cover fraction is a significant source of uncertainty for L-band InSAR retrievals of SWE, but this technique may be the first space-borne technique capable of retrieving SWE below non-dense forest canopy without any a priori information.