Land-surface emissivity maps based on MSG/SEVIRI information

Transcription

1 Land-surface emissivity maps based on MSG/SEVIRI information Leonardo F. Peres, Carlos C. DaCamara Centro de Geofísica da Universidade de Lisboa (CGUL) and Instituto de Ciência Aplicada e Tecnologia (ICAT), Faculty of Sciences, University of Lisbon, Campo Grande, Lisbon, Portugal ldperes@fc.ul.pt; ccamara@fc.ul.pt ABSTRACT Retrieval of land-surface temperature (LST) using data from MSG/SEVIRI requires adequate estimates of land-surface emissivity (LSE). In this context LSE maps for SEVIRI channels IR3.9, IR8.7, IR10.8, and IR12.0 were developed based on the vegetation cover method. A broadband LSE map (3-14 µm) was also developed for estimating long-wave surface fluxes that may prove to be useful in both energy balance and climate modelling studies. LSE is estimated from conventional static classifications on land cover, LSE spectral data for each type of land cover, and fractional vegetation cover (FVC) information. Both IGBP-DIS and MODIS MOD12Q1 land cover products were used to build the LSE maps. Data on LSE were obtained from the JHU and JPL spectral libraries included in the ASTER spectral library, as well as from the MODIS-UCSB spectral library. FVC data for each pixel were derived based on NDVI. Depending on land cover type the LSE errors for channels IR3.9 and IR8.7 spatially vary from ± 0.6% to ± 24% and ± 0.1% to ± 33%, respectively, whereas the broadband spectrum errors lie between ± 0.3% and ± 7%. In the case of channels IR10.8 and IR12.0, 73% of the land-surfaces within the MSG disk present relative errors less than ± 1.5% and almost all (26%) of the remaining areas have relative errors of ± 2.0%. Developed LSE maps provide a first estimate of the ranges of LSE in SEVIRI channels for each surface type and obtained results may be used to assess the sensitivity of algorithms where an a priori knowledge of LSE is required. 1. INTRODUCTION An accurate retrieval of land-surface temperature (LST) based on remote sensing measurements from space requires a proper characterization of the atmospheric influence as well as an adequate knowledge of land-surface emissivity (LSE). In this context, we have built LSE maps for SEVIRI channels IR3.9, IR8.7, IR10.8 and IR12.0 in order to provide information on emissivity for LST retrievals from MSG/SEVIRI data. Our main purpose is to provide the required information on LSE by the split-window (SW) technique since it is the procedure currently being used to derive LST on an operational basis in the framework of the Satellite Application Facility on Land Surface Analysis (LSA-SAF). In this respect, LSE maps will also allow identifying those areas where SW may be successfully applied. Besides, LSE maps are expected to provide first-guess estimates to be used in algorithms based on the two-temperature method (TTM) that allow a simultaneous retrieval of LST and LSE [1], [2]. Finally, a broadband LSE map (3-14 µm) is also developed for estimating long-wave surface fluxes that may prove to be useful in both energy balance and climate modelling studies where constant values of LSE are usually assumed because of the limited knowledge about its spatial variation [3]-[5]. 1

2 2. METHOD AND DATA The computation of LSE maps relies on the so-called Vegetation Cover Method (VCM) [6] where land-surface is considered as an heterogeneous (both in LST and LSE) and rough system, that is made of a mixture of vegetation and ground where the row crops form a cavity consisting of crop walls with ground in between. The effective emissivity ε c of such system in a given channel c is given by [7] ε c = ε c, d + dε c (1) where the first term in the right hand side of (1) is related to the radiance directly emitted by the system. Assuming that the vegetation top and side have the same emissivity ε c,v, then the term ε c, d becomes ε c, d = ε c,v FVC + ε c, g (1 FVC ) (2) where ε c, g is the emissivity of the ground and FVC denotes the fractional vegetation cover. The second term in the right hand side of (1) is related to the radiance indirectly emitted through internal reflections occurring between crop walls and the ground, and is given by [ ] dε c = (1 ε c, g )ε c,v F (1 FVC ) + (1 ε c,v )ε c, g G + (1 ε c,v )ε c,v F ps (3) where ps is the fractional amount of the vegetation side, and F, G, and F are geometrical factors as defined in [7] and may be determined if the height H of the different boxes and the distance S between them are known. Since the theoretical expression of dε c is quite complex, we have followed [8] and adopted a quadratic form to simulate the factor dε c. Accordingly, (3) becomes dε c = 4 dε c FVC (1 FVC ) (4) where dε c is defined in this work by the mean of different values of dε c that are computed for reasonable ranges of H and S. Data on LSE were obtained from the Johns Hopkins University (JHU) and Jet Propulsion Laboratory (JPL) spectral libraries included in the ASTER spectral library, as well as from the MODIS-UCSB spectral library. Based on the definition of channel LSE, ε c (θ ), the spectral LSE values were used to compute LSE within SEVIRI channels IR3.9, IR8.7, IR10.9, and IR12.0 as follows λ =λ2 f (λ )ε (λ, θ ) B(λ, T )dλ c ε c (θ ) = λ = λ1 s (5) λ = λ2 f (λ )B(λ, T )dλ λ λ c = s 1 2

3 where f c (λ ) is the spectral response function of the sensor in channel c, B(λ, Ts ) denotes the emitted radiance given by Planck s function for the surface temperature Ts, and λ1 and λ 2 are respectively the lower and upper limits of the channel spectral range. The broadband LSE is defined in a similar way. In order to compute the LSE maps we have used a global land cover database, referenced hereafter as IGBP-DIS, generated by the U.S. Geological Survey s (USGS) Earth Resources Observation System (EROS) Data Center, the University of Nebraska-Lincoln (UNL) and the Joint Research Centre of the European Commission (JCR) [9]. Among the different available land cover classifications we have adopted the IGBP Land Cover Classification [10], which distinguishes 17 different classes: 1Evergreen Needleleaf Forest; 2- Evergreen Broadleaf Forest; 3- Deciduous Needleleaf Forest; 4- Deciduous Broadleaf Forest; 5- Mixed Forest; 6- Closed Shrublands; 7- Open Shrublands; 8- Woody Savannas; 9- Savannas; 10Grasslands; 11- Permanent Wetlands; 12- Croplands; 13- Urban and Built Up; 14Cropland/Natural Vegetation; 15- Snow and Ice; 16- Barren or Sparsely Vegetated; 17- Water. Alternatively, we have also used the MODIS MOD12Q1 land cover product [11], which provides a suite of land cover classifications with the primary classification according to the IGBP scheme. This product is computed quarterly allowing the assessment of yearly changes in land cover. Land cover maps covering the MSG full disk were generated within the framework of the LSA-SAF Project [12] and are based on the normalized geostationary projection [13]. Both IGBP-DIS land cover and MODIS MOD12Q1 land cover product for MSG full disk are shown in Fig. 1. Fig. 1. Land cover maps for the MSG full disk and the IGBP land cover classes; (a) MOD12Q1 (2000/01) and (b) IGBP-DIS (1992/93) land cover products. A simple way to estimate FVC is by means of the normalized difference vegetation index (NDVI). Considering a mixed pixel of vegetation and ground, and reflectance measurements ρ in the near infrared channel VIS0.8 and in the red channel VIS0.6 we may write the NDVI value as NDVI = FVC ( ρ 0.8,v ρ 0.6,v ) + (1 FVC ) ( ρ 0.8, g ρ 0.6, g ) FVC ( ρ 0.8,v + ρ 0.6,v ) + (1 FVC ) ( ρ 0.8, g + ρ 0.6, g ) (6) 3

4 where v and g denotes respectively the vegetation and ground. The fraction of vegetation cover is obtained by inverting (6) 1 FVC = NDVI NDVI g ρ ρ0.6,v 1 NDVI 0.8,v 1 NDVI NDVI ρ g 0.8, g ρ0.6, g NDVIv (7) where NDVI, NDVI g and NDVI v are respectively the vegetation indices for the mixed pixel, the ground and the vegetation surfaces. NDVI and FVC maps for the MSG disk (see Fig. 2) were then computed based on reflectance synthetic images (SEVIRI channels VIS0.6 and VIS0.8) generated within the framework of the LSASAF Project [12]. Fig. 2. Maps of NDVI (a) and FVC (b) for the MSG full disk. The maps are based on reflectance synthetic images for SEVIRI channels VIS0.6 and VIS0.8. Water bodies and cloud fields are shown in black. Since the used spectral libraries only provide reflectance measurements for individual samples, we have first applied thermal infrared (TIR) versions of bidirectional reflectance distribution function (BRDF) models to relate component reflectance measurements to those of structured surfaces [14], [15]. Accordingly, the row crops LSE in VCM were computed from the BRDF volumetric model [16], whereas a rough-surface specular BRDF model [14] was used in the case of water and ice. The assignment of LSE from direct measurements (ground) and from the volumetric (vegetation) and specular (water and ice) models to IGBP classes was performed based on about 150 samples that were used individually or combined in order to characterize the 17 IGBP land cover classes. Combination of vegetation and ground samples and specification of the respective weight were based on the description of each class, as given in [10]. 3. ANALYSIS AND RESULTS 4

5 Based on the combination of vegetation and ground samples used to describe the 17 IGBP classes, we have characterized LSE in channels IR3.9, IR8.7, IR10.8, IR12.0, and in the broadband 3-14 µm by means of their weighted mean values and the maximum deviations from the mean values. These deviations are defined as the largest absolute difference between the mean and the two extreme values of the LSE class range. It must be emphasized that values of deviations provide the information on the maximum LSE error due to the LSE variability in each class. The mean LSE values for channels IR3.9, IR8.7, IR10.8, and IR12.0 range respectively from to 0.996, to 0.997, to 0.997, and to whereas for the broadband spectrum the range is between and For most classes the lowest mean LSE values are found in channel IR3.9 whereas the greatest ones occur in channel IR12.0. The same behaviour may be observed on the LSE variability where the values for channel IR3.9 present the largest fluctuation. In general, the results also show that the LSE values for vegetated areas are greater and present less variability than those from areas without vegetation. In order to determine the value of dε c, a reasonable range of H and S was defined based on the description of each class [17]. We have also performed a sensitivity analysis on VCM in order to assess the total LSE error for each class. Defining the limiting error on LSE for channel c by the 1norm, we have δε c = FVC δε c,v + 1 FVC δε c, g + ε c,v ε c, g + 4 dε c (1 2 FVC ) δfvc + 4 FVC (1 FVC ) δ dε c (8) Typical values of FVC were then assigned for each IGBP class according to the surface type description; the maximum LSE deviations were then used to characterize the LSE errors of vegetated and uncovered areas (i.e., δε c,v and δε c, g ); and results from [17] (see Table IX) were finally used to define the uncertainty on dε c. We have also used relative errors on FVC varying between ± 5% and ± 25%, which is an appropriate range for FVC estimates based on vegetation indices [6]. Fig. 3 shows the relative errors on LSE of each IGBP class for channels IR3.9, IR8.7, IR10.8, IR12.0, and broadband spectrum. Based on the considered range of uncertainties in FVC, we may observe that the LSE error varies slowly with the FVC error. This behaviour of VCM is explained by the fact that the term associated with the FVC error corresponds to the difference between vegetated and ground LSE, which is a small quantity since in general both LSE values are similar [6]. It is worth noting that only classes 16 (Barren or Sparsely Vegetated) and 15 (Snow and Ice) present relative errors greater than ± 1.5% for channels IR10.8 and IR12.0. This is explained by the fact that class 16 is composed of different materials and therefore it is the most difficult class to typify, thus presenting a larger LSE variation. On the other hand, the two components (snow and ice) of class 15 present different behaviour, the snow being Lambertian at all wavelengths and ice being predominantly specular in TIR. In fact, as snow grains become larger the snow approaches the spectral and directional behaviour of ice resulting into a strong specular component in the TIR region [18]. However, our 5

6 results show that the difference between the LSE values of ice and fine snow (as well as the difference between coarse snow and fine snow) is only significant for channel IR12.0 explaining the larger LSE error in this band. Although large variations of LSE may be observed in desert areas (class 16), the broadband map may provide valuable information as many energy balance and climate modelling studies have assumed a constant and uniform LSE [3]-[5]. Fig. 3. LSE relative errors for channels IR3.9 (a), IR8.7 (b), IR10.8 (c), IR12.0 (d), and for broadband spectrum (e) with respect to errors in FVC ranging from ± 5% to ± 25%. It is worth noting that the LSE error for each land cover class must be taken into account together with the respective proportion of the total land-surface area [15]. Results shown in Fig. 3 and the computed percentage of pixels in the MSG disk that belong to each class (for details see [17]) indicate that about 73% of the landsurfaces within the MSG disk present relative errors less than ± 1.5% for channels IR10.8 and IR12.0 and almost all (i.e., 26%) of the remaining areas, corresponding to class 16 (Barren or Sparsely Vegetated), have relative errors of ± 2.0%. As a final result of the VCM application, we present in Fig. 4 LSE maps for the MSG full disk respecting to the channels IR3.9, IR8.7, IR10.8, IR12.0 and broadband spectrum. These maps are based on both the MODIS MOD12Q1 land cover product and the FVC parameter derived as described in Section CONCLUSIONS Obtained results have shown that the LSE errors in the obtained map for IR3.9 range from ± 0.6% to ± 24% and that such variability is too high when compared with channels IR10.8 and IR12.0 (from ± 0.1% to ± 2.4%). Thus, we believe that when used together with the developed LSE maps, a SW algorithm only including channels 6

7 IR10.8 and IR12.0 will provide more accurate LST estimates. In the case of channels IR10.8 and IR12.0, approximately 73% of the land-surfaces within the MSG disk present relative errors less than ± 1.5% and almost all (26%) of the remaining areas, corresponding to class 16 (Barren or Sparsely Vegetated), have relative errors of ± 2.0%. LSE errors for the broadband spectrum vary between ± 0.3% and ± 7%. Developed maps provide a first estimate of the ranges and variations of LSE in SEVIRI channels IR3.9, IR8.7, IR 10.8 and IR 12.0 for each surface type and therefore, obtained results may be further used to assess the sensitivity of algorithms where an a priori knowledge of LSE is required. In what respects to the broadband LSE map (3-14 µm), it will allow taking into account the spatial and spectral variations of LSE in energy balance and climate modelling studies where a constant and uniform LSE is often used [3]-[5]. Developed LSE maps may be viewed as the base line to derive more refined and accurate LSE maps. Validation of LSE maps is currently being planned and will rely on the comparison of the LSA-SAF products with those based on other in orbit instruments such as MODIS and ASTER. Finally, it is worth mentioning that developed LSE maps are currently being used within the framework of the LSA-SAF Project, which has already started its initial operational phase. Fig. 4. LSE maps corresponding to MSG full disk based on MODIS MOD12Q1 land cover product and FVC information. (a) Channel IR3.9, (b) channel IR8.7, (c) channel IR10.8, (d) channel IR12.0, and (e) broadband LSE. Water bodies and cloud fields are represented in black color. REFERENCES 7

8 [1] L. F. Peres, and C. C. DaCamara, Inverse problem theory and application: Analysis of the twotemperature method for land-surface temperature and emissivity estimation, IEEE Geoscience and Remote Sensing Letters, vol. 1, pp , [2] L. F. Peres, and C. C. DaCamara, Land surface temperature and emissivity estimation based on the Two-Temperature Method: Sensitivity analysis using simulated MSG/SEVIRI data, Remote Sensing of Environment, vol. 91, pp , [3] J. L. Kinter, J. Shukla, L. Marx, and E. K. Schneider, A simulation of the winter and summer circulations with the NMC global spectral model, Journal of the Atmospheric Sciences, vol. 45, pp , [4] NMC Development Division, Documentation of the research version of the NMC medium-range forecasting model, NMC Develop. Div., Natl. Meteorol. Cent., Camp Springs, MD, 504 pp., [5] X. Zeng, M. Shaikh, Y. Dai, R. E Dickinson, and R. B. Myneni, Coupling of the common land model to the NCAR community climate model, Journal of Climate, vol. 14, pp , [6] V. Caselles, E. Valor, C. Coll, and E. Rubio, Thermal band selection for the PRISM instrument 1. Analysis of emissivity-temperature separation algorithms, Journal of Geophysical Research, vol. 102, pp , [7] V. Caselles, and J. A. Sobrino, Determination of frosts in orange groves from NOAA-9 AVHRR data, Remote Sensing of Environment, vol. 29, pp , [8] E. Valor, and V. Caselles, Mapping land surface emissivity from NDVI: Application to European, African, and South American areas, Remote Sensing of Environment, vol. 52, pp , [9] T. R. Loveland, B. C. Reed, J. F. Brown, D. O. Ohlen, J. Zhu, L. Yang, and J. W. Merchant, Development of a global land cover characteristics database and IGBP DIScover from 1-km AVHRR data, International Journal of Remote Sensing, vol. 21, pp , [10] A. S. Belward, and T. Loveland, The IGBP-DIS 1km land cover project, remote sensing in action, in Proc. 21st Annu. Conf. Remote Sensing Society, Southampton, UK Sept., 1995, pp [11] M. A. Friedl, D. K. McIver, J. C. F. Hodges, X. Zhang, D. Muchoney, A. H. Strahler, C. E. Woodcock, S. Gopal, A. Schnieder, A. Cooper, A. Baccini, F. Gao, and C. Schaaf, Global land cover from MODIS: Algorithms and early results, Remote Sensing of Environment, vol. 83, pp , [12] LSA-SAF Project Team, Scientific Prototyping Report, EUMETSAT internal documentation, Doc. No. SAF/LAND/SPR/1.0, Issue 1.0, [13] R. Wolf, Coordination Group for Meteorological Satellites LRIT/HRIT Global Specification, Doc. No. CGMS03, Issue 2.6, [14] W. C. Snyder, and Z. Wan, BRDF models to predict spectral reflectance and emissivity in the thermal infrared, IEEE Transactions on Geoscience and Remote Sensing, vol. 36, pp , [15] W. C. Snyder, Z. Wan, Y. Zhang, and Y.-Z. Feng, Classification-based emissivity for land surface temperature measurement from space, International Journal of Remote Sensing, vol. 19, pp , 1998 [16] J. Roujean, M. Leroy, and P. Deschamps, A bi-directional reflectance model of the eath s surface for correction of remote sensing data, Journal of Geophysical Research, vol. 97, pp , [17] L. F. Peres, and C. C. DaCamara, Emissivity maps to retrieve land-surface temperature from MSG/SEVIRI, IEEE Transactions on Geoscience and Remote Sensing, vol. 43, pp , [18] J. W. Salisbury, D. M. D Aria, and A. Wald, Measurements of thermal infrared spectral reflectance of frost, snow, and ice, Journal of Geophysical Research, vol. 99, pp ,