¶¶Òõ¶ÌÊÓƵ

Abstract: Satellites provide the most consistent and reliable measurements of snow and ice globally with estimates of snow-covered area, grain size, and concentration of light-absorbing particles creating global potential to improve existing operational forecasts and enable forecasting of water supplies for regions with little to no in situ forecasting. A main hurdle to consistent, accurate estimates of snow properties from space is the accurate and automatic identification of clouds in satellite images. Clouds cover snow from the view of sensors in the visible and infrared and impact the energy balance of the snowpack as it accumulates and melts.While most clouds are brighter in shortwave-infrared wavelengths than snow, some mixed pixels of snow and soil or vegetation can have multispectral signatures that are similar to those for some thin clouds.  Some measured snow and cloud spectra can be successfully modeled as both snow and cloud spectra with no ability to discriminate, based on the common differences between snow and cloud, in the broad band passes of existing moderate resolution satellites. These spectral similarities limit the performance of cloud masks that class pixels individually. Fortunately, both snowpack’s and clouds are far larger than the spatial resolution of the multispectral satellite data we use to measure them, so an alternative approach is to use superpixels (segmented regions of local similarity) as the spatial resolution for data processing. Superpixels can automatically segment images into groups of pixels based on regional low-level image characteristics and replace pixels as the spatial resolution for processing; this greatly reduces the image size but not the information. Superpixels easily incorporate spatial and textural features alongside spectra for accurate cloud identification and show promise in improving cloud masks over snow covered terrain. The benefits of superpixels extend past the cloud masking phase and can be used as the spatial resolution much farther into an image processing workflow to increase speed and accuracy of algorithms designed for multispectral and hyperspectral sensors. 

This talk will give an overview of why we need better forecasts of water supply that use satellite data, the current limitations of operational cloud masks over snow covered mountainous regions, recent advances in global cloud masking algorithms, and some of our work on superpixels we are incorporating into our workflows for masking clouds, measuring fractional snow cover, and estimating snowmelt. Various approaches to creating superpixels and their feature sets will be explained along with how they can be used throughout the imaging processing workflow for reducing imaging noise, speeding up algorithms, and training machine learning models.  

Timbo Stillinger, PhD
Postdoctoral Scholar
Earth Research Institute, UC Santa Barbara