16
Development and evaluation of Development and evaluation of Passive Microwave SWE Passive Microwave SWE retrieval equations for retrieval equations for mountainous area mountainous area Naoki Mizukami

Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

Embed Size (px)

Citation preview

Page 1: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

Development and evaluation of Development and evaluation of Passive Microwave SWE Passive Microwave SWE retrieval equations for retrieval equations for

mountainous areamountainous area

Naoki Mizukami

Page 2: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

1. 1. IntroductionIntroductionWith increasing researches into remote sensing application to snow hydrology, it has been shown the space-borne passive microwave data has potential capability of extracting snowpack volume - snow water equivalent (SWE) & snow depth (SD) - for global and regional scales. Remote sensing could provide spatially distributed snowpack data for any area of interest where few ground observations are made.

The objective of this study is to examine the correlations between passive microwave data from Special Sensor Microwave Imager (SSM/I) and observed SWE for mountainous regions.

Page 3: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

2. 2. BackgroundBackgroundPassive microwave SWE retrievals have been developed empirically for global and regional scales.

For global scale SD mapping, Chang et al. (1987) first applied a single form of empirical equation – SD = a x BTD with constant a where BTD is brightness temperature difference between 2 channels. Kelly et al. (2003) improved this algorithm by implementing spatially varying coefficient of a, reduced RMSE by 4.0 ~ 8.0 cm.

Page 4: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

For regional scale, for instance, Meteorological Service of Canada (e.g. Goodison and Walker, 1995) have been involved into development/improvement of regional SWE retrievals for the prairie region in eastern Canada.

For mountainous regions, however, few studies on passive microwave SWE retrievals have been conducted in detail probably because of 1) sparse snow observations for calibration/validation of the algorithm, 2) heterogeneous spatial snowpack distribution compared to satellite spatial resolution.

Page 5: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

3. 3. Data setData set Daily Snowpack Telemetry (SNOTEL) SWE Daily SSM/I brightness temperature (Tb)

7 channels (19, 37 and 85GHz) with dual polarizations (vertical and horizontal) .

The pixel size is 25 km x 25km

Period – 11 winter seasons (1992-93 through 2002-03), 6 months for each winter season (November through April).

Page 6: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

4.4. Analysis ProcedureAnalysis Procedure1. Used a single SSM/I pixel that encompass

several SOTEL sites in the Wasatch Mountains.

2. Aggregate daily SWE time series and daily Tb time series into 10 day values.

3. Spatially aggregate 10 day SWE from multiple SNOTEL sites at the center of SSM/I pixel using inverse distance weighting

4. Perform stepwise regression (forward selection) to develop linear regression equations in monthly basis. The t-stat was used for stopping rule (max p-value = 0.05).

5. Model evaluation – RMSE and bias and linear correlation between microwave based SWE and SNOTEL SWE, and sensitivity test.

Page 7: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

Predictors (21 candidates)

Predictand = observed SWE

19H-19V

19V-37H

37H-37V

37V-85H

85H-85V

19H-37H

19V-37V

37H-85H

37V-85V

19H-37V

19V-85H

37H-85V

19H-85V

19V-85V

19H-85V

Single Tb: 19H, 19V, 37H, 37V, 85H, and 85V

Brightness temperature difference (TBD)

Polarization difference is considered to remove effect of snow wetness on microwave.

TBD is considered to remove effect on snowpack temperature on microwave

Page 8: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

5. 5. Study sitesStudy sites

-113 -112.5 -112 -111.5 -111 -110.5 -11040

40.5

41

41.5SNOTEL sitesSSM/I pixel center

Longitude

lati

tud

e

Fig 1. The SSM/I pixel (orange dot) and 4 SNOTEL sites (yellow dots) within 25km from the center of the pixel used for regression analysis

UTWY

Page 9: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

6. 6. Regression resultsRegression results

Regression equation

Nov

SWE=-1.72+0.33(19V-85V)-0.25(37H-85V)

Dec SWE=-8.33-0.16(19H-85V)+0.84(19H-37H)-0.73(85H-85V)

Jan SWE=179.73-0.66(19H)

Feb SWE=175.17+0.67(19H)

Mar

SWE=2.29+2.41(19V-85H)-2.20(19H-85V)

Apr SWE=194.57+0.73(37H)

Table 1. Resultant monthly regression equations. Maximum P-value for t-statistic for the predictors to be added was 0.05. When no predictors less than 0.05 of p-value were left, stepwise regression analysis stopped.

Page 10: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

7. 7. Predicted versus Observed SWEPredicted versus Observed SWE

Fig 2. Time series of SWE estimated by monthly regression equations and SWE based on SNOTEL observations (top panel) and residual (lower panel)

0

5

10

15

20

25

30Microwave SWE

SNOTEL SWE

1991 1993 1994 1995 1997 1998 2000 2001 2002 2004-10

-5

0

5

10

SW

E,

cmR

esid

ual,

cm

Time

Page 11: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

7. 7. Predicted versus Observed SWEPredicted versus Observed SWE

0 5 10 15 20 250

5

10

15

20

25NovDecJanFebMarApr

Fig 4. Scatter plot of estimated SWE from empirical equations versus SWE from SNOTEL

SNOTEL SWE, cm

Mic

row

ave d

eri

ved

SW

E,

cm

R

Nov 0.81

Dec 0.86

Jan 0.66

Feb 0.46

Mar 0.63

Apr 0.74

overall

0.87

Page 12: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

Nov Dec Jan Feb Mar Apr0

0.2

0.4

0.6

0.8

1

Underestimate

Unbiased

Sta

nd

ard

ized

R

MS

E

Month

7. 7. Error characteristicsError characteristics

Fig3. RMSE and bias for each month. The bias between -0.1 and 0.1 is plotted as “unbiased” . During spring, the equations tend to underestimate and RMSE . RMSE was standardized by standard deviation of predictand (SWE).

Page 13: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

Fig 5. Frequency of R2 from regression analysis with one winter season excluded from 11 winter seasons in total (11 R2

in total). The regression analysis used the same forms of equation as the ones from stepwise regression (table 1)

8. 8. Sensitivity testSensitivity test

R2

Cou

nt

0

1

2

3

4

5Nov Dec Jan

0 0.5 10

1

2

3

4

5Feb

0 0.5 1

Mar

0 0.5 1

Apr

Page 14: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

9. 9. Model evaluation at different Model evaluation at different pixelpixel

-113 -112.5 -112 -111.5 -111 -110.5 -11040

40.5

41

41.5SNOTEL sitesSSM/I pixel center

Longitude

lati

tud

e

Fig 6. The SSM/I pixel (orange dot) and 5 SNOTEL sites (yellow dots) within 25km from the center of the pixel used for independent model evaluation. The pixel enclosed by orange circle was used for regression analysis (see Fig. 1)

UTWY

Page 15: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

9. 9. Model evaluation at different Model evaluation at different pixelpixel

0 5 10 15 20 25 30 35 400

5

10

15

20

25

30

35

40NovDecJanFebMarApr

SNOTEL SWE, cm

Mic

row

ave d

eri

ved

SW

E,

cm

Fig 6. Scatter plot of estimated SWE from empirical equations (table 1) versus SWE from SNOTEL. Obvious underestimate during spring.

Underestimate during spring

R

Nov 0.71

Dec 0.80

Jan 0.61

Feb 0.31

Mar 0.34

Apr 0.63

overall

0.82

Page 16: Development and evaluation of Passive Microwave SWE retrieval equations for mountainous area Naoki Mizukami

10. 10. Summary and Future studySummary and Future study Multiple regression equations were

developed using SNOTEL SWE data and passive microwave Tb in monthly basis.

Monthly regression equations explain 0.7-0.8 of total variability of SWE overall.

There is some difficulty in estimating SWE in February –March (underestimate).

Future studies include Gridding ground-based SWE with statistical

interpolation such as kriging. Test other equation types (polynomial) for

regression analysis Test the empirical equations in the different

mountain regions