Transcript Estimating Snow Depth in Complex Terrain

Estimating Snow Depth in
Complex Terrain
Andy Newman (RAL/HAP)
ASP Research Review
17 July 2015
Outline
•
•
•
•
•
What?
Where?
Why?
How (and how difficult)?
Research topics
▫ Preliminary results
▫ Future work
What and Where?
• Estimate snow depth or snow water equivalent
(SWE)
• Mountainous areas above tree line
Why?
• Many regions depend on
snow melt for a large
portion of their water
needs (Western US)
▫ Irrigation
▫ Drinking water
▫ Hydroelectricity
Why?
• Snow cover strongly influences the surface
energy budget, land surface variables (i.e.
sensible flux, soil moisture, vegetation)
• Areas of deep snow stick around longer
▫ Impacts intensity and timing of spring melt
Liston et al. 1995
Why?
• 25% (or more) of mountain
snowpack above tree line
• Extreme wind
redistribution
▫ Snow depth extremely nonuniform
▫ Net transport to below tree
line
▫ Wind & sublimation losses
very uncertain
Images courtesy of Ethan Gutmann
How?
• How does one estimate snow depth or SWE?
▫ Measurements or model estimates
• Many ways to measure snow depth/SWE
How?
• Statistical and time-integrated models
▫ Statistical models provide a snapshot of snow depth or
SWE
 Maximum depth prior to melt onset
 Use terrain based parameters (no snow)
▫ Elevation, slope, aspect, etc.
 Perform well in some localized test areas
▫ Column or spatially-distributed snow depth evolution
models
 Distributed models
▫ 1 - m to many km grid spacing (ideally 1-10 m)
▫ Account for wind transport, sublimation,
precipitation/melting, snowpack morphology
How?
• Example scene for distributed model
Liston et al. 2007
How Difficult?
• Measurements
▫ Point measurements
 How representative is one point?
 Challenging locations
▫ Distributed and areal measurements
 Snow surveys and LIDAR flights are expensive, satellite
estimates lack resolution
• Models
▫ Statistical models
 Can they generalize to large areas?
▫ Distributed models




How to distribute sparse observations?
Quality terrain and vegetation datasets?
Model equations correct?
Computational concerns
Research Topics
• Can we advance current statistical and/or
distributed models to further improve snow pack
estimates?
▫ Off-line distributed simulations forced by highresolution mesoscale model output
 Verification using LIDAR observations
 Examining potential model improvements
▫ Artificial neural networks (ANNs) may provide a
more robust statistical method
 How robust are they?
Research Topics
• ANNs use inter-connected nodes
(neurons) to perform complex
pattern recognition, regression
• Input layer connected to neuron
layer
▫ Non-linear transfer functions
• Neurons connected to output
layer
• Very good for non-linear
statistical modeling with many
input parameters
Wikipedia
Preliminary Results
• Initial ANN:
▫ Inputs
 Terrain based parameters
▫ Elevation, slope, aspect, curvature, sheltering index
Wind Dir.
Winstral et al. 2002
Preliminary Results
• ANN trained and validated on 1 million
randomly selected points (snow depth)
▫ 70% for training, 30% for validation and testing
▫ Explains ~42% of total variance in snow depth
▫ Comparable to results of Winstral et al. (2002)
Preliminary Results
• ANN estimated snow-depth
▫ Underestimates very deep snow, more shallow snow cover
▫ Distribution of ANN snow depth too Gaussian
Future Work
• Investigate non-local terrain parameters for
ANN
• Validate ANN over wide array of locations
• Determine distributed model performance and
areas needing improvement
Questions?