Dw. Cline et Tr. Carroll, Inference of snow cover beneath obscuring clouds using optical remote sensing and a distributed snow energy and mass balance model, J GEO RES-A, 104(D16), 1999, pp. 19631-19644
We describe a new method to infer the presence or absence of snow cover whe
n it is obscured to remote sensing by clouds. The inference is based on the
spatial distribution of snow water equivalent (SWE) estimated by a physica
l snow model. The snow model accounts for surface and internal snowpack ene
rgy exchanges and mass exchanges including snow accumulation, sublimation,
rainfall, and meltwater outflux. We examine two approaches to using the mod
eled SWE information to infer snow cover: (1) we directly categorize areas
with modeled SWE greater than zero as "snow covered" and areas with zero SW
E as "no snow"; (2) we classify the modeled SWE field into a binary map of
snow covered/no snow, using maximum likelihood logistic estimation (LR). He
re the modeled SWE serves as a semicontinuous independent variable, and the
binary dependent variable consists of observed snow cover derived from all
available ground observations of snow depth and water equivalent, and from
samples of snow cover randomly selected from cloud-free areas of a satelli
te snow cover product. We demonstrate these methods for a 504,000 km(2) reg
ion in the north central United States, over a 4 week period during March a
nd April 1997. The snow model was run hourly on a 4 km grid using data norm
ally available in near real time, including numerical weather analysis prod
ucts, satellite-derived insolation products, and ground observations of pre
cipitation. We tested two dependent variable data configurations for the LR
approach to simulate (1) typical conditions when remotely sensed snow cove
r observations are available to help develop the logistic model and (2) "wo
rst-case" conditions where only ground-based data are available. Averaged o
ver the study period, all three methods, the direct SWE, the worst-case LR,
and the typical LR, yielded comparable results in the range of 78-80% accu
racy when compared to satellite-observed snow cover maps. The results of th
is study demonstrate that a relatively simple, spatially distributed, physi
cally based snow model is capable of providing useful snow cover informatio
n in an operational environment.