Plot-scale retrieval of land surface temperature (LST) and emissivity (ε)) estimation
Supported by the Luxembourg National Research Fund (FNR) ATTRACT programme (WAVE, A16/SR/11254288)
Gitanjali Thakur, Stan Schymanski, Kaniska Mallick, Ivonne Trebs Luxembourg Institute of Science and Technology
Conclusions
INTRODUCTION METHODOLOGY RESULTS
MOTIVATION BACKGROUND
PROBLEM
RESEARCH QUESTIONS
LST ESTIMATION ε ESTIMATION
UNCERTAINTY
PLOT-SCALE LST PLOT-SCALE ε LST COMPARISON LST UNCERTAINTY
2
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
SURFACE ENERGY BALANCE AND SURFACE TEMPERATURE:
2
Radiometer
H (W m
-2) is sensible heat, LE (W m
-2) is latent heat, G (W m
-2) is ground heat flux, σ is Stephan-
Boltzmann constant and ε is the surface emissivity, T
s(K) is surface temperature, T
a(K)
is air temperature, K is the conductance to heat transport between surface and atmosphere.
Eddy covariance system
H LE G
●
Surface energy balance (R
net) depends on surface temperature (T
s)
H, T
aare
measured
●
K is an important parameter determining surface
atmosphere coupling T
sis
required
to derive K
3
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
RADIATIVE ENERGY BALANCE AND SURFACE TEMPERATURE:
3
Radiometer
R
net(W m
-2) is net radiation, R
ldwnis down-welling longwave, R
lrefis reflected longwave, R
srefis reflected shortwave, R
lemis emitted longwave, R
lupis up-welling longwave, σ is the Stephan-Boltzmann constant and ε is the surface emissivity and T is surface temperature
Eddy covariance system
R
sdwn+ R
ldwnR
sref+ R
lrefR
lemR
lup●
Net radiation (R
net) is difference between the down- welling and up-welling radiative components
●
Radiometer measures up-welling longwave ( i.e emitted longwave + reflected longwave)
●
Land surface temperature at plot-scale is calculated by inverting R
lupequation
4
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
PLOT-SCALE LST ESTIMATION : INCONSIS -TENCY IN EQUATION
4
Simplified equation (seq) Complete equation (leq)
Assuming ε ≈ 1
●
In past R
ldwnwas not measured routinely at Eddy covariance sites, resulting omission of reflected longwave component, and simplification of equation.
Even now with the availability of R
ldwn ,use
of simplified equation for LST estimation is a
common practice
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
LST AND EMISSIVITY AT PLOT SCALE :
RQ 1. Is simplified equation adequate to estimate LST at plot-scale?
RQ 2. How to obtain correct emissivity needed for plot-scale LST retrievals using tower-based measurements?
RQ 3. What is the resulting uncertainty in diurnal LST due to measurement uncertainties using complete and simplified equation?
6
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
LAND COVER TYPES:
●
Ten sites
[1][2]having good record of eddy covariance data
Site Name Land cover types Adelaide River (AR) Savanna dominated by
Eucalyptus Alice Spring (AS) Mulga Canopy Daly uncleared (DU) Woodland savana Howard Spring (HS) Woodland Savanna Litchfield ( LF) Tropical Savanna
Sturt Plains (SP) Grassland (Mitchell grass) Ti Tree East (TT) Grassy mulga woodland &
Triodia savanna Tumbarumba (TUM) Wet Sclerophyll forest Brookings (BR) Cropland
Yatir Forest (YF) Evergreen needle forest
1) http://data.ozflux.org.au/portal/home.jspx 2) https://ameriflux.lbl.gov/
Table 1: Study sites for the analysis
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
METHODS : RESEARCH QUESTION 1
Plot scale LST using complete and simplified equation?
Step 1: MODIS spectral emissivity for Channel 31 & 32 provides landscape-scale broadband emissivity
Step 2: Daytime longwave measurement and landscape-scale emissivity (ε
MODIS) is used to calculate plot-scale LST using complete equation (T
leq) and simplified equation (T
seq).
8
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
Fig. 1 : Distributions of diurnal surface temperature for a representative year at each study site. Surface temperatures based on short equation (Tseq ) and complete equation (Tleq) using landscape-scale emissivity (εMODIS). εMODIS is derived using spectral emissivity from channel 31 & 32. The median values of Ts are shown at top of the plot and the emissivity used for the Ts retrieval are shown at bottom in orange.
ε
MODISPLOT SCALE LST : USING LEQ AND SEQ
How different are T
seq, T
lequsing landscape scale emissivity (ε
MODIS) ?
Simplified equation leads to higher values than complete equation
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
SENSITIVITY OF T
SEQAND T
LEQTO EMISSIVITY:
Fig. 2: An illustration of surface temperature (Ts ) sensitivity to emissivity using complete and simplified equation.
Longwave measurement (Rlup, Rldwn) used for the plot is from Alice Springs on 15/06/2018 at midday
How sensitivity of T
sto emissivity differs using complete and simplified equation?
Tleq is less sensitive to emissivity used
10
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
Linear regression between H And Ts – Ta should pass through
origin if estimated Ts is correct
H=0 when T
s−T
a= 0
Step 1: Segregating each month daytime data (H, R
ldw, R
lup, T
a) Step 2: Assuming ε range between 0.4 to 0.998
Step 3: Calculate T
sfor each value of emissivity
Step 4: Linear regression between sensible heat (H) and (T
s-T
a) is forced through origin and coefficient of determination (R
2) and root mean square error (RMSE) is calculated
Step 5: If R
2> 0.5, select ε resulting in lowest RMSE
Sensible heat is driven by surface-air temperature difference (Ts - Ta )
METHODS: RESEARCH QUESTION 2
Theory
EXAMPLE
H= K (T
s−T
a)
How to estimate plot-scale emissivity using H and T
s- T
a?
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
PLOT-SCALE EMISSIVITY: USING SEQ AND LEQ
Lower sensitivity of Tleq results in lower value of ε
Fig. 3: Sensible heat (H) vs ∆T (Ts − Ta ) regression based on the simplified and complete equation is presented for August 2005 at Brookings. (a) Reproduction of analysis presented in Fig. 2 (a) in Holmes et al. (2009) based on simplified equation. (b) Replication of Fig. 2 (a) using complete equation. Blue crosses mark data points satisfying the filtering criteria while black dots mark points not considered in the analysis. N is the number of blue crosses used for regression (red line), m is the slope of regression, RMSE is the root mean square error and R2 is the coefficient of determination. The fitted ε value is reported in the title.
How T
seqand T
leqresults into different plot-scale emissivity?
12
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
COMPARISON: PLOT-SCALE AND MODIS EMISSIVITY
Site Name MODIS emissivity Plot-scale emissivity
SEQ LEQ
Adelaide River (AR) 0.985 0.99 0.96
Alice Spring (AS) 0.974 0.96 0.82
Daly uncleared (DU) 0.985 0.99 0.98
Howard Spring (HS) 0.985 0.92 0.60
Litchfield ( LF) 0.985 0.92 0.60
Sturt Plains (SP) 0.974 0.96 0.85
Ti Tree East (TT) 0.974 0.95 0.80
Tumbarumba (TUM) 0.983 0.99 0.97
Brookings (BR) 0.983 0.98 0.82
Yatir Forest (YF) 0.974 0.97 0.83
Table 2: MODIS emissivity (εMODIS ) are obtained using channel 31 and 32 . The plot-scale emissivity represented above are median of monthly emissivity obtained analysing three years of H vs Ts – Ta plots for Australian site and one year data for Brookings and Yatir.
How different are plot-scale emissivity to the landscape-scale emissivity (MODIS) ?
Complete equation has lower sensitivity resulting into lower emissivity
13
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
LST COMPARISON:
How plot-scale T
scalculated using ε
MODIS, ε
optcompares to MODIS LST ?
Fig. 4: Comparison of plot-scale LST (Tseq , Tleq ) with landscape-scale LST (TMODIS ) derived from daily MODIS overpass. (a) Tseq based on short equation and MODIS derived landscape-scale emissivity; (b) Same as (a), but Tleq based on complete equation; (c) Tseq based on short equation and monthly plot-scale emissivity; (d) Same as (c), but Tleq based on complete equation. Bias is mean of Tplot scale − T MODIS , N is the number of daily overpasses of MODIS between 2016 and 2018. c is the intercept, m the slope, RMSE is the root mean square error and R2 is coefficient of determination
Use of plot-scale emissivity
reduces the bias between landscape and plot-scale LST
14
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
How do errors in measured fluxes result in ε and Ts uncertainty?
Step 1 : Select uncertainty bounds for each input variable (H, T
a,R
lup,R
ldwn)
Step 2 : Uniformly distributed samples are generated using Saltelli sample generator Step 3 : Optimum emissivity range for each sample (measured + error sample) is generated
Step 4: Uncertainty in hourly LST resulting due to uncertainty in ε is estimated
H bound = [-20, 20]
Rlup bound = [-5, 5]
Rldwn bound = [-5, 5]
Ta bound = [-1, 1]
Uniformly distributed samples
generated within bounds
Measured (H , Rlup , Rldwn , Ta) +
generated samples
Uncertainty in
Hourly LST
Range of optimum emissivity for each month
UNCERTAINTY: PLOT-SCALE LST
Saltelli sample generator
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
UNCERTAINTY: DIURNAL LST
Fig.6: Uncertainty in hourly Ts − T a for 15 August 2017. (a) Hourly daytime uncertainty in Tleq − Ta due to perturbed fluxes and optimum ε using complete equation in blue and perturbed fluxes with MODIS based ε in orange. (b) Hourly daytime uncertainty in Tseq − Ta due to perturbed fluxes and optimum ε using simplified equation in black and perturbed fluxes with MODIS based ε in orange. Unperturbed optimum ε values and T s − T a values correspond to the median of perturbed values. Uncertainty in ε
Uncertainty in hourly T s − T a reduced when plot-scale ε is
used
Comparison of uncertainty in hourly T
s- T
ausing plot-scale and landscape-scale ε ?
Additional input (Rldwn ) in complete equation is not increasing Ts - Ta uncertainty
16
Previous Home Next
1. Introduction 1.1 Motivation 1.2 Background
1.3 Problem
1.4 Research questions 1.5 Study sites
2. Plot-scale LST 2.1 Methods 2.2 Result
3.Plot Scale emissivity 3.1 Holmes et al. (2009) 3.2 Result
4.LST uncertainty 4.1 Method 4.2 Result 5. Conclusions
●
Simplified equation produces different results (LST, ε) to complete equation and therefore should not be used
●
The use of plot-scale emissivity for LST retrieval reduces the bias between tower based in-situ LST and landscape-scale (MODIS) LST
●
The uncertainty in T
s-T
ais reduced using plot-scale emissivity in comparison to MODIS based constant emissivity
Results
Results
Results
CONCLUSIONS:
Previous Home
1) Holmes, T. R. H., De Jeu, R. A. M., Owe, M., & Dolman, A. J. (2009). Land surface temperature from Ka band (37 GHz) passive microwave observations. Journal of Geophysical Research: Atmospheres, 114(D4)
2)Holmes, T. R., Hain, C. R., Anderson, M. C., & Crow, W. T. (2016). Cloud tolerance of remote-sensing technologies to measure land surface temperature. Hydrology and Earth System Sciences, 20(8), 3263-3275
REFERENCES :
ACKNOWLEDGMENTS:
We would like to thank Dr. Maik Renner for pointing us to the work by Holmes et al. and Dan Yakir’s lab for providing Yatir Forest data and helpful discussions. We are also grateful to
Thomas Foken, Jason Beringer, Lindsay Hutley and Mauro Sulis for insightful discussions.This
work is supported by the Luxembourg National Research Fund (FNR) ATTRACT programme
(WAVE, A16/SR/11254288).
18
Calculation of Ts using
different epsilon values from the range defined. (Step 3)
RMSE values calculated for the corresponding epsilon values. (Step 4)
Optimized value of epsilon giving the least RMSE value. (Step 5)
Fig. A: H vs ΔT (T
leq-T
a) plots illustrating the steps for obtaining optimized emissivity
EXAMPLE ESTIMATION OF EMISSIVITY
To method
RANGE OF UNCERTAINTY FOR EMISSIVITY
Seq resulted in more constrained values
between 0.94 and 0.9
Fig. A1: Uncertainty in plot scale monthly emissivity for 2017 due to perturbation in H, R
lup, R
ldw, T
aat using complete and simplified equation shown in Blue and black at Alice Springs
To Result