ESTIMATION OF AIR AND SURFACE TEMPERATURE EVOLUTION OF THE EAST ANTARCTIC SHEET BY MEANS OF PASSIVE MICROWAVE REMOTE SENSING M. Brogioni , G. Macelloni, S. Pettinato, F.Montomoli IFAC - Institute of Applied Physics National Research Council Firenze, Italia International Geoscience and Remote Sensing Symposium Vancouver, Canada, 24-29 July, 2011 /20
Antarctica is the coldest and emptiest place on Earth
Antarctica influence directly the Earth climate due to its extension (14-30 million of km 2 ) and average temperature ~ -50°C
As a comparison: Arctic 8 million of km 2 , Greenland 2 million of km 2 , Europe 10 million of km 2
It is one of the most important indicators of the climate changes
Knowledge about Antartica is limited due to the harsh environment
Antarctica Monagham, WWI Mag. 22.1 /20 South Pole Antarctic Peninsula 4% of Antarctica (like California) Glacial retreats are widespreads and move to South West Antarctica 20% of Antarctica (like Greenland) Stores 6m of global sea level Marine based (it rests over the sea) It is shrinking overall East Antarctica 76% of Antarctica (larger than USA) Stores 60m of global sea level Approximatively in balance Mean altitude ~3000m
4/20 Aim of the work Passive microwave sensors are working since the 80s’ and can image Antarctica several times per day (up to 8 in the Dome C region (75° S) Antartica is the most undersampled continent due to the cost of the manned exploration and the difficulties related to the impervious environment The use of remote sensing techniques can help in monitoring the spatial and temporal characteristics of large regions. Some interesting topics are the spatial and temporal evolution of temperatures , the snow mass balance, the detection of melting zones.
MW and Snow temperature data (Dome-C) The temporal behavior of Tb was closely related to the snow temperature at different depths
This analysis was conducted on more than 25000 images (at least five images per day)
The mean value of the 3x3 pixel area was extracted from each image in order to reduce noise.
T 50 37 GHz T 100 19 GHz
/17 37 GHz – 10 cm 19 GHz – 200 cm Correlation analysis Examples of correlation between snow temperature and brightness Determination coefficient (R 2 ) between Tb and Snow Temperature at different depths Frequency T50 T100 T200 T300 T400 T600 T800 T1000 6.9 GHz 0.63 0.72 0.62 0.38 0.13 0.08 0.62 0.54 10 GHz 0.74 0.87 0.78 0.51 0.19 0.07 0.73 0.68 19 GHz 0.83 0.94 0.80 0.50 0.17 0.10 0.80 0.70 37 GHz 0.98 0.90 0.55 0.19 0.01 0.39 0.83 0.43
Experimental data /20 AMSR-E : More than 45000 images Frequencies used: Ku, Ka V polarization Time: January 2003- December 2008 AWS snow and air temperature measurements: GREEN - Australian Antarctic Survey BROWN - University of Wisconsin* PURPLE - Italian National Project for Researches in Antarctica* *Dome C data were collected also during the IFAC Domex experiment
AWS sites Sites of the AWS considered in this work /20 AGO 1 AGO 4 Panda S AGO 3 AGO 5 Dome C GC 41 Giulia Irene Dome A Eagle Panda N LGB 46 LGB 35 LGB 20 Dome Fuji Relay Station Mizuho JASE 2007 West Antarctica Peninsula East Antarctica No data were available in the period 2003-2008 Only air temperature was available Air and snow temperature available
Methodology /20 The study was carried out by using linear regressions between ground measurements and satellite data (i.e. Tair and Tb 37GHz, Tsnow 50cm and Tb 19 GHz). In order to keep the temporal variability of the datasets, Tair, Tsnow and Tb were not temporal averaged. Here we considered up to 8 measurements per day. Brightness temperature were spatially averaged over a 3x3 pixel area in order to lower the noise. This has a tiny impact since the std dev of the 9 measurements is lower than 1K. We didn’t use ANN techniques (already considered in previous works) because their performances seems to be comparable to the ones of the regressions for this kind of study.
Snow temperature retrieval
/17 Snow temperature retrieval (Dome C) ANN REGRESSIONS Developed for the year 2005 Developed for the year 2006
/20 Snow temperature retrieval (Dome A, Eagle) Dome A and Eagle ground data were obtained from Australian AWS In these sites, AWS measured Tsnow at 0.1, 0.3, 3 and 10 m below the surface only Tsnow at 1m is estimated in this work. Eagle (76.43°S, 77.02°E) Dome A (80.44°S, 77.21°E) Regressions between T b 19GHz and T snow 1m R 2 0.89 R 2 0.95
/20 Dome A (80.44°S, 77.21°E) Eagle (76.43°S, 77.02°E)
We verified that (at least in these sites) it is possible to estimate snow temperature 1m below the surface with an RMSE of about 1.5K,
The error seems to be stable throughout the years
Test of the method was carried out at different latitude and longitude
Results of snow temperature retrieval
It is worth noticing that:
The accuracy of the retrieval (i.e. the RMSE) and the determination coefficients (R 2 ) obtained, makes this study useful for estimating the snow sub-superficial temperature when precision of 1K are sufficient,
Snow temperature retrieval
For climatological issues, the obtained precision could not be adequate (i.e. if the accuracy required is one order of magnitude higher),
It seems somewhat difficult to lower the RMSE of the relationships since the accuracy of the measuring instruments (i.e. the AMSR-E and SSM/I radiometers) is of the same order (around 1.5K).
/17 Air temperature retrieval
/17 Snow temperature variations are primarily driven by air temperature fluctuation which heat (and cool) the snow by convection. This is different from land surfaces whose temperature depends on the solar radiation. Tair and Tsnow (on which depends the microwave Tb) are quite good correlated, making possible an attempt to estimate air temperature from Tb measurements. Correlation between air and Tb at 37GHz (the highest frequency commonly used in the remote sensing of snow) is not high as with the Tsnow due to the heat latency of snow. Few remarks Data collected at Eagle in 2005 In order to obtain better performances it is useful to consider the temporal changes of snow emissivity
/17 Snow equivalent emissivity In order to perform the Air temperature retrieval by means of MW data, we used an equivalent emissivity of snow obtained as because the snowpack is subject to metamorphic changes due to the weather conditions (mainly air temperature and wind action). Eagle Dome A
/17 Eagle Dome A LGB35 Examples of air temperature retrieval results Usually the average regression provide the best results!
19/20 Results of the air temperature retrieval Despite the quite high R 2 , the mean RMSE obtained is not very good (betw.4 and 8K)
Possible causes can be:
- the heat latency of snow which damped the Tair variations, making the Tb slightly "insensitive" to the Tair variations,
- the quality of the AWS data were not always good due to the enviromental conditions which in some cases affect the normal service of the AWSs
Site Average relationship R 2 Years considered Mean RMSE (°C) 2003 2004 2005 2006 2007 2008 Eagle Tb 37GHz = 0.5256 T air + 222.82 0.746 5.65 Dome A Tb 37GHz = 0.5486 T air + 211.75 0.775 8.06 LGB20 Tb 37GHz = 0.6180 T air + 233.64 0.843 5.4 LGB35 Tb 37GHz = 0.8280 T air + 226.40 0.919 3.9 Dome Fuji Tb 37GHz = 0.6207 T air + 216.25 0.744 7.48 Mizuho Tb 37GHz = 0.7334 T air + 203.91 0.62 6.28 Relay station Tb 37GHz = 0.5642 T air + 223.42 0.757 6.43 Giulia Tb 37GHz = 0.7059 T air + 221.93 0.854 5.25 Irene Tb 37GHz = 0.4918 T air + 233.14 0.692 8.71
/20 Future works
The results found outline that it is possible to retrieve the snow and air temperature from microwave data, albeit with a RMSE error of some degrees.
Next steps of this work will be the exploitation of the spatial and temporal trends of the retrieved Snow and Air temperatures over a long time period (since the 80’s) in order to assess the climate variations on the East Plateau. This will be obtained by using passive microwave data, consolidated relationships between Tsnow, Tair and Tb, and assimilation methods (like kriging).
Future possible improvements could be obtained by the joint use of microwave and infrared images, albeit these latter are affected by the weather conditions and, for a certain extent, by the diurnal cycle.
/20 Thanks for the attention!
Why to study Antarctica for the climate changes
Experimental data description
Retrieval of snow temperatures
/20 Penetration depth (1/e) Model Analysis:Contribution of Layers (0-100 m) Multilayer model based on the Strong Fluctuation Theory Input: experimental data from Epica and Domex campaigns 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Depth (m) Layer Contribution (%) C X Ku Ka
Retrieval of snow temperature : 1 – 10 meters /17 Δ T = Maximum – Minimum Temperature, R 2 = Correlation coefficient , SE = Standard Error of Estimate, Err = Mean Percentage Error T100 T200 T300 T400 T500 T600 T800 T1000 ΔT [°C] 25 17.03 10.30 6.74 4.31 3.39 1.60 0.99 R 2 0.95 0.89 0.90 0.89 0.94 0.97 0.90 0.88 SE [°C] 1.9 1.21 0.74 0.54 0.26 0.17 0.12 0.08 SE/ ΔT [%] 7.6 7.1 7.2 8 6 5 7.5 8.1 Very good correlation !
/17 Retrieval of snow temperature : 1 m 100 cm y = 0.9909x + 0.5475 R 2 = 0.9531 100 cm Trained 2005 Retrieved 2008 y = 1.0679x + 2.9725 R 2 = 0.9688 Trained 2005 Retrieved 2006
Previous study: retrieval of Tsnow 0-2 meters /17 Data measured for the year 2006 compared with the retrieved one. Relationship between Tb and Tsnow for the year 2005 were used for the retrieval R 2 =0.98, SE=1.5 °C R 2 =0.95, SE= 1.9 °C
The electromagnetic model /17
The Brightness Temperature Tb was computed according to the wave approach which accounts for reflection and transmission between the layers by means of the propagating matrix (Kong,1990).
The V and H components of Tb were obtained by adding the contributions of the snow layers by means of the fluctuation dissipation theorem (Jin,1984).
The obtained value of Tb was the results of the average of 50 realizations each one corresponding to a profile of ( z )
Model input parameters: /17
Density ( z ) was modeled as:
( z ) = m + f ( z ) ;
m = measured mean value; f = fluctuating part
< f ( z 1 ) f ( z 2 )> = p 2 exp (- z 1 – z 2 / l z ) (Gaussian)
The correlation length was obtained from a semi-empirical relationship derived from ice core data permittivity was computed from the strong fluctuation theory as a function of correlation length and density
Snow Temperature and Grain Sizes were obtained from measurements
The snow measurements /17
Model Analysis : Contribution of Layers (0-100 m) /17 Penetration depth (1/e) 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Depth (m) Layer Contribution (%) L C X Ku Ka
/17 The Getis local statistic of the i-th pixel sum of the weight in the window x and s are the mean and the standard deviation of the entire image x j value of the j-th image pixel weight of the pixel : 1 if the pixel belong to the window, 0 elsewhere Spatial and temporal analysis (I) i j
/17 Spatial and temporal analysis Analysis performed on 2 orbits (22 and 23 images) in 2008 0 0.5 1 1.5 2 2.5 3 Std Dev (K) 6.8 GHz 4 5 6 7 8 9 10 11 12 13 Std Dev (K) 37 GHz
/17 Spatial and temporal analysis (II) 2 orbits (22 and 23 images) in 2008 0 0.5 1 1.5 2 2.5 3 Std Dev (K) 6.8 GHz 4 5 6 7 8 9 10 11 12 13 Std Dev (K) 37 GHz Maps Isolines Temporal Std Dev Spatial Getis statistic
/17 6.8 GHz Spatial and temporal analysis (II) Dome C Maps Isolines Temporal Std Dev Spatial Getis statistic 0 0.5 1 1.5 2 2.5 3 Std Dev (K)
/17 Based on the previous study, we performed a regression analysis in order to retrieve the snow temperature Algorithm developed by using data collected in 2005 snow temperature retrieved for the year 2006 Snow temperature retrieval RMSE=1.16K RMSE=1.64K
/17 The algorithm was tested also with the Tsnow data of year 2008 Similar analysis were performed by developing algorithms for the years 2006, then validating them with data collected in different years. Then, the retrieval was performed also by using ANN in a feed-forward multi-layer perceptron scheme (MLP) with some hidden layers of neurons between the input and output. Snow temperature retrieval RMSE=1.01K RMSE=1. 43K
/17 Although it is not possible to verify the retrieved snow temperature values, these considerations indicate that the Tsnow estimation do not present appreciable problems There is always a delay between the Tair and Tsnow temperature. The range of T100 values is lower than the T50 one, which is in turn lower than the air temperature swing. It is also worth noticing that the maximum in the Tair (which happened in 2002) corresponds to the maximum of the estimated Tsnow. Retrieval of Tsnow for the past years Sebbene nn ci siano dati per verifica
/17 Analysis of temperature trends The trend in the air temperature shows an increase of 1.3°C in the period 1997-2008 A first analysis seems to confirm that the temperature of the first layers increases Can the emissivity constantly increase? Why Tb are constantly increasing?