UC Santa Barbara
Estimating Snowmelt in an Alpine Environment Using Satellite Resources
- Author(s): Kamal-Heikman, Shithi
- Advisor(s): Dunne, Thomas
- et al.
Snowmelt contribution to streamflow in the western United States is a critical supply of fresh water. The most widely applied operational model for estimating snowmelt utilizes the temperature index model which is based on empiricism and ground data. This work aims to advance our capability for estimating snowmelt for both clear and cloudy weather conditions using a ground calibrated physically based modified radiation balance that utilizes satellite data. The robustness of a modified radiation balance for computing snowmelt depends on reliable estimates of both long and short-wave radiation. I tested my model estimations of snowmelt against daily measurements of snow water equivalent (SWE) reduction at snow pillows from 129 locations over a 14-year period in the snowmelt dominated region of the Sierra Nevada.
For clear-sky conditions (n ~ 25,000), I found that the temperature index model achieved an R2 = 0.52 (RMSE =9.7 mm/day;) with remotely sensed LST as an input and R2=0.48 (RMSE = 10.1 mm/day) using measured air temperature (Ta) from weather stations. The modified radiation balance model achieved an R2 = 0.72 (RMSE = 7.4 mm/day) with a fixed atmospheric emissivity, while a further refinement of an emissivity calculated from satellite-derived precipitable water resulted in R2 =0.75 (RMSE = 7.0 mm/day).
In order to estimate shortwave radiation during cloudy days, I first constructed a simple binary cloud index based on inference of clouds from ‘missing’ MODIS LST data, which when regressed against daily shortwave ratio, yielded a regression equation with a correlation coefficient of R2 = 0.40 and an RMSE = 0.21. For an improved method, I used cloud optical depth (COT) and cloud mask (MOD35) data from MODIS 06C6 and found that with an exponential decay equation the regression with shortwave ratio led to a correlation coefficient R2 = 0.72 and a lower RMSE=0.14 making this method the more robust the two.
For estimating the long-wave radiation, a closer examination of data on directly measured longwave radiation at one individual station revealed that measured air temperature is a better predictor of long wave compared to the satellite based LST. I found that the relation between daytime Ta and MODIS LST, as well as the diurnal LST swing, was influenced by surface characteristics such as canopy cover. I used the findings to introduce an LST “swing factor” into a linear equation for modeling daytime Ta based om LST. Utilizing Ta modeled from LST the long-wave radiation model performance improved significantly compared to a model based directly on LST. I extended my snowmelt model to include cloudy-days based on temporal extrapolation of LST data and an emissivity cloud-adjustment based on remote sensed cloud properties.
With the cloud adjustments for both short- and long-wave incorporated in the snowmelt model, I added several physical factors over the course of 11 model runs. These included the effect of canopy in the local environment and a simple albedo model, that can cause spatial and temporal variations in the radiation load. The addition of cloudy days increased the data sample size significantly (n=58,077), with the final model run achieving an R2 = 0.68 with an RMSE= 7.0 mm/day. I achieved my goal of increasing the physical basis of the model for estimating snowmelt in both clear- and cloudy sky conditions, and for modeling the radiative effect of clouds without compromising model capability for estimating snowmelt at the snow sensor scale. In the course of my research, I was also able to identify several promising paths for future work for improving the physical basis and accuracy of the model.