Machine Learning on Images: Combining Passive Microwave and Optical Data to Estimate Snow Water Equivalent in Afghanistan’s Hindu Kush

Jeff DozierUniversity of California, Santa Barbara

University of Washington, 2016-02-25

News item from IRIN (UN)http://www.irinnews.org/Report/93781/Analysis-Afghan-drought-conditions-could-spell-disaster

KABUL, 21 September 2011 (IRIN) – “The current dry spell sweeping across Afghanistan’s northern, northeastern and western provinces could lead to a large-scale food crisis and the humanitarian community should act quickly to ensure this does not degenerate into a disaster, government and aid officials warn.”

AMSR2,1.49 km3

Reconstruction,3.54 km3

SNODAS,5.22 km3

10 100

200

300

400

500

600

700

800

900

103Snow water equivalent, Sierra Nevada, mm, 2014-04-01

Question: can we estimate the real snow water equivalent, in places like the Hindu Kush, from information available during the season?

2011-04-01, passive microwave SWE, 27 km3

SWE, mm snow fraction

2011-04-01, MODIS snow cover, 190,000 km22011-04-01, reconstructed SWE, 68 km3

Fractional snow cover from MODIS2015-04-01

𝑅𝜆=𝜖𝜆+∑𝑘=1

𝑁

𝑓 𝑘𝑅 𝜆 ,𝑘 Data from http://snow.jpl.nasa.gov

Comparison of MODIS (500m) and Landsat (30m) snow fraction, in the Sierra Nevada

32 scenes with coincident MODIS and Landsat imagesAverage RMSE = 7.8%Range from 2% to 12%

Energy balance reconstruction

American River basin Snow Pillows

on  day𝑛 ,𝑆𝑊𝐸𝑛=𝑆𝑊𝐸0−∑𝑗=1

𝑛

𝑀 𝑗

when𝑆𝑊𝐸𝑛=0 ,𝑆𝑊𝐸0=∑𝑗=1

𝑛

𝑀 𝑗

𝑀 𝑗=𝑀𝑝 𝑗× 𝑓 𝑆𝐶𝐴 𝑗

From energy balance model, driven by NLDAS or GLDAS

𝑀𝑝 𝑗

11

Downscaling

Sola

r rad

iatio

nLo

ngwa

ve ra

diat

ion

Water balance? Over a year so is ? (Only for Reconstruction)

Comparison of reconstructed SWE with Airborne Snow Observatory

Comparison of reconstructed SWE with Airborne Snow Observatory

Sierra Nevada, AMSR-E & Reconstruction

Persistent sources of uncertainty in passive microwave retrieval of SWE:deep snow, vegetation, sub-grid heterogeneity

[Luojus et al., GlobSnow ATBD, http://www.globsnow.info/docs/GS2_SWE_ATBD.pdf]

Question: can we estimate the real snow water equivalent from information available during the season?

2011-04-01, passive microwave SWE, 27 km3

SWE, mm snow fraction

2011-04-01, MODIS snow cover, 190,000 km22011-04-01, reconstructed SWE, 68 km3

The problem• Discover a pattern that uses real-time data—passive

microwave SWE and snow-covered area from optical sensor—to match reconstructed SWE, which is available only after the snow is gone

Approach• Among the choices for machine learning methods, start

with neural networks• Investigate alternatives later

• Develop a training set with 2003-2004 data, and predict 2005, from March 1 to June 1• Then adapt 2005 results, and predict 2006• And so on

Results

(modest improvement in R2 over time, but unpredictable bias)

2009 integrated total SWE (km3)

Insight from comparing statistical distributions (red predicted, blue reconstructed)

Improvements to the reconstruction

Remotely sensed albedo of fractional snow

Dirty snow albedo has a similar spectral shape to fine-grain clean snow

[Warren, Rev Geophys, 1982]

Errors in assimilated energy inputs

“Clouds” in cloud-free scenes (Landsat OLI, Sierra Nevada)a

2013-05-20 2015-03-07

Steps forward (but already better than what we have now)• Improve reconstruction

• Sensitive to albedo, so fix the albedo estimate of fractionally snow-covered pixels

• Consider alternatives to NLDAS/GLDAS• Solve the snow-cloud discrimination

• Incorporate better microwave snow retrieval methods• Improve machine learning

• Examine other ensemble statistical methods, e.g. genetic programming, regression boosted decision trees, support vector machines

• Find another input variable that helps with the bias, e.g. satellite precipitation products, models like WRF

• But . . . We’re encouraged by the results, as a way to estimate snow water equivalent in mountains without existing sensors

“the author of all books” (James Joyce, Finnegan’s Wake)

Slides available at http://www.slideshare.net/JeffDozier/

Finis

Data available at ftp://ftp.snow.ucsb.edu/pub/org/snow/users/dozier/MachineLearning/