| Summary: | Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Oceanographic Engineering at the Massachusetts Institute of Technology and the Woods Hole Oceanographic Institution September 2020. Sea ice thickness has long been an under-measured quantity, even in the satellite era. The snow surface elevation, which is far easier to measure, cannot be directly converted into sea ice thickness estimates without knowledge or assumption of what proportion of the snow surface consists of snow and ice. We do not fully understand how snow is distributed upon sea ice, in particular around areas with surface deformation. Here, we show that deep learning methods can be used to directly predict snow depth, as well as sea ice thickness, from measurements of surface topography obtained from laser altimetry. We also show that snow surfaces can be texturally distinguished, and that texturally-similar segments have similar snow depths. This can be used to predict snow depth at both local (sub-kilometer) and satellite (25 km) scales with much lower error and bias, and with greater ability to distinguish inter-annual and regional variability than current methods using linear regressions. We find that sea ice thickness can be estimated to ∼20% error at the kilometer scale. The success of deep learning methods to predict snow depth and sea ice thickness suggests that such methods may be also applied to temporally/spatially larger datasets like ICESat-2. This research was funded by National Aeronautics and Space Administration grant numbers NNX15AC69G and 80NSSC20K0972, the US National Science Foundation grant numbers ANT-1341513, ANT-1341606, ANT-1142075 and ANT-1341717, and the WHOI Academic Programs Office.
|
|---|