Comparison of NDVI derived from NOAA / AVHRR LAC and PAL data

Comparison of NDVI derived from NOAA/AVHRR LAC and PAL data Martin F. Garbulsky1,3 and José M. Paruelo 2,3. 1 Depart. de Producción Animal. Facultad de Agronomía. Universidad de Buenos Aires. Av. San Martín 4453. C1417DSE. Buenos Aires. Argentina. 2 Depart. de Recursos Naturales y Ambiente. Facultad de Agronomía. Universidad de Buenos Aires. 3 Laboratorio de Análisis Regional y Teledetección. IFEVA. Tel: 54-11-4524-8070 Fax: 54-11-4514-8730 garbulsky@ifeva.edu.ar. Keywords: NOAA/AVHRR, Local Area Coverage , LAC, Pathfinder, PAL, NDVI, vegetation. Abstract - NOAA/AVHRR satellite imagery has been widely used for vegetation studies at global and regional scales, to study for example the exchange of matter and energy between the ecosystem and the atmosphere or to generate land cover maps. PAL and LAC products are usually processed as maximum composite yielding an image every 10 days, with 8 and 1 km spatial resolution, respectively. However differences in the processing algorithms and on the data available to generate the products, radiometric data and spectral indices derived are not equal. The objective of this analysis was to generate mathematical relationships between data provided by LAC and PAL imagery, in order to make them comparables for vegetation studies. We work with 36 NDVI 10 day composites for the 1992-93 growing season from each of the databases for southern South America. The LAC images were degraded to the same spatial resolution of the PAL images (8km) in order to make both sources comparables. We worked on the values of individual pixels and on the annual NDVI average (NDVII). Our results showed that the relationship between NDVI-I from both sources differed from de 1:1 line: NDVI-I LAC = 0.77 x NDVI-I PAL + 0.092 (r2=0.82). We detect non-random spatial differences that can be partially associated to differences on the processing. Our results suggest that a direct relationship between NDVI sources is useful in order to make the sources compatible for vegetation studies. I. INTRODUCTION The use of NOAA/AVHRR satellites images for ecosystem functioning studies had an increasing importance since the ’80. Global (Loveland et al, 2000), continental (De Fries et al., 1998) and local (Duchemin et al., 1999) vegetation studies had used the Normalized Difference Vegetation Index provided by those satellites for the analysis of the exchange of matter and energy between the vegetation and the atmosphere and to generate land use maps. Four main products were available for this kind of studies: Global Vegetation Index (GVI), Global Area Coverage (GAC), Local Area Coverage (LAC) and Pathfinder AVHRR Land project (PAL). The first product of the program, the GVI database (Tarpley et al. 1984), comprises images for 34 consecutives months between 1982 and 1984 with 15 km spatial resolution. The GAC database was built on the base of the same AVHRR data but with 4 km spatial resolution (http://edcdaac.usgs.gov/1KM/avhrr_sensor.html). LAC database images have 1 km spatial resolution and 10day temporal resolution (Eidenshink and Faundeen, 1994). Lastly, PAL database has 20 year data, the longest AVHRR database, with 8 km spatial resolution and 10 day temporal resolution temporal (James and Kalluri 1994). These two former products are the most widely used nowadays (e. g.: Pelkey et al., 2000; Di Bella et al, 2000, Loveland et al, 2000). Simultaneous utilization of NOAA/AVHRR imagery with different spatial resolution processing for vegetation studies o their combination with data coming from other platforms (e.g. Landsat, SAC-C, Modis), requires a proper correspondence between data sources. User’s manuals and published bibliography do not alert on the existence of differences on NDVI values among images with different spatial resolution (Agbu and James, 1994). Different NDVI processing arrived to divergent values, however scarce bibliography was published on the comparison of the NDVI patterns generated form different data sources (Young and Anyamba 1999). The objective of this analysis was to analyze the correspondence and to generate mathematical relations between data provided by NOAA/AVHRR LAC and PAL in order to make them comparables in southern South America. II. METHODOLOGY We used 36 images (10 day composites) from each of the Local Area Coverage (LAC) and Pathfinder AVHRR Land (PAL) program for the period July 1992 – June 1993, a whole growing season for the southern hemisphere (20 – 56° S.; 48 – 73° W. aprox.). These images come form the same original data, but different processing was done to arrive to spatial resolution of 1km for LAC and 8 km for PAL images. PAL images are obtained from the Global Area Coverage (GAC) images. The GAC processing was done in the beginning of the NOAA/AVHRR program due to the low storage capacity on board. The GAC processing consists in the averaging of 4 contiguous pixels, resuming the information in new pixels of 1.1 by 4 km of spatial resolution (http://edcdaac.usgs.gov/1KM/avhrr_sensor.html). PAL NDVI are obtained on the base of the maximum NDVI value from GAC pixel to obtain 8x8 pixels (Agbu and James, 1994). We produced a new dataset to make LAC and PAL databases comparables, by the spatial degradation of the LAC images. These images were spatially degraded by averaging NDVI values from 64 contiguous pixels (1x1) for each of the 36 composites. Hence, we obtained 36 NDVI images coming from the LAC database with 8 km spatial resolution, hereafter degraded LAC (LACdeg). We extracted NDVI data from PAL and LACdeg images by a 10x10 pixels grid sampling all over the studied region (figure 1). We discard sea-masked pixels, analyzing then 800 pixels for each composite. We calculated the NDVI annual average (NDVI-I) for each group of images. We calculated linear regression between NDVI and NDVI-I values of each database. III. RESULTS Spatial distribution of the NDVI-I differences between data sources was not random neither. Pixels with positive values in NDVI-I LACdeg - NDVI-I PAL differences (figure 1) were concentrated mainly in Atacama Desert, Patagonia and along the Andes (Puna and continental glaciers in southern Argentina and Chile). Areas with differences on the other sense are concentrated in southern Brazil and with less importance in the southern Andes and along the main rivers. Differences between data sources are not uniform along the NDVI gradient, as is evident from the difference in slope from 1 and the difference in the intercept from 0. For low NDVI values, PAL images provide a lower NDVI value than LAC. An inverted situation arose for values above 0.55. Difference between both average NDVI was 0.007. 0.8 a) NDVI 0.6 0.4 0.2 0 Jul-92 Nov-92 Mar-93 NDVI-I Nov-92 Mar-93 NDVI-I 0.8 b) NDVI 0.6 0.4 0.2 0 Jul-92 0.8 c) 0.6 NDVI Atlantic Ocean 0.4 0.2 0 Jul-92 Nov-92 Mar-93 NDVI-I 0.8 d) NDVI 0.6 0.4 0.2 0 Jul-92 Nov-92 Mar-93 NDVI-I Figure 1. Extent of the analysis of the PAL and LAC images and spatial pattern of NDVI-I differences. Black areas denote pixels where NDVI-I PAL-NDVI-I LACdeg differences were higher than 0.05. In white, pixels with negative differences lower than –0.05. Grey areas denote areas with near zero differences. Right plots show the NDVI seasonal patterns and the NDVI-I in four areas a) Atacama desert, b) Southern Brazil, c) Argentinean Pampas and d) Northern Patagonia. ¡ are NDVI LACdeg values and Î are NDVI PAL. 1 a) Slope 0.8 value. The temporal variability on the coefficient of determination (fig. 2) shows that there are certain composites along the analyzed period, with a high dispersion on the data from both data sources. 0.8 NDVI-I LACdeg The form of the areas with differences in NDVI-I has particular features: areas south of 37° with low NDVI values (Patagonia shrub steppes) presented a pattern related with the satellite orbit (fig. 1). Straight differences in southern Brazil presented characteristics related with the composite processing. Composite #33 (June 93, fig. 1b) presented NDVI difference around 1.5. This difference could explain the difference in NDVI-I for that area. White areas in Atacama Desert do not presented an associated satellite orbit pattern. 0.6 0.4 0.2 0.6 0.4 0.0 0.0 0.2 0.2 0.4 b) Intercept 0.3 0.2 0.6 0.8 NDVI-I PAL 0 Jul-92 Sep-92 Nov-92 Ene-93 Mar-93 May-93 0.4 Figure 3. Relationship between the annual average NDVI (NDVII) for PAL and LACdeg images. NDVI-I LACdeg = 0.7786 x NDVI-I PAL + 0.09236 r²=0.8258; p<0.0001; n=698. Relationship between NDVI-I calculated from PAL and LAC presented a high coefficient of determination: r2=0.825 (fig. 3). The regression line for that relationship was: 0.1 Jul-92 Sep-92 Nov-92 Ene-93 Mar-93 May-93 NDVI-I LACdeg = 0.778 x NDVI-I PAL + 0.09236 The slope and the intercept of this regression were statistically different from the 1:1 line. 1 IV. DISCUSSION 2 c) Coef. of correlation (r ) 0 0.8 0.6 0.4 0.2 0 Jul-92 Sep-92 Nov-92 Ene-93 Mar-93 May-93 Figure 2. Seasonal variability in the regression parameters for the relationship between PAL and LACdeg NDVI. a) Slope, b) intercept and c) coefficient of correlation for 36 dates. Straight lines show the annual average for each parameter. All the regressions were significant p<0.0001; 540<n<807. The parameters of the relationship between NDVI values of each source (slope, intercept and coefficient of determination) presented temporal variability along the analyzed period (fig. 2). The average of the slopes for the 36 ten-day composite was 0.704, 0.117 for the intercept and 0.594 for the coefficient of determination. Minimum coefficient of determination was 0.274 for the third composite of September and corresponds with a low slope Regression between NDVI-I and NDVI for different spatial resolution images showed that different processing drove to significant differences for low NDVI areas. Negative differences (PAL<LAC) found south of 37° parallel should be related with the composting of winter images, because of the orbital forms of the strips in figure 1. Secondly, uniform differences along the year are no related with particular composites. Differences in those areas could be originated in a deficient light incidence in high latitudes particularly in winter and to the lower chance of cloudless pixels. Atmospheric and zenithal angle corrections, especially at high latitudes, could generate diverging results. The magnitude of the differences found shows that there are problems with the indistinctly use of both data sources for vegetation studies. For example, taking into account the NDVI-I in northern Patagonia, the calculated value with LACdeg data is 0.17, meanwhile 0.08 with PAL data (fig. 1d). Using the equation presented by Paruelo et al. (1997) to estimate Annual Net Primary Productivity, we observed important differences: LACdeg would throw 138 g C.m-2.y-1 while 37 g C.m-2.y-1 in the case of PAL data. Differences found in southern Brazil in particular composite images and with less importance in the Argentinean pampas, are of less impact for vegetation studies. Anomalous data of this kind is feasible of detection by means of visual inspection or simple algorithms that detect and discard that data. For example composite 33 in LACdeg data set should be discard for southern Brazil (fig. 1b). One of PAL processing errors already detected is associated to the seasonal variation in the solar zenith angle. This difference, although insignificant, was detected by NASA and images were reprocessed (ftp://daac.gsfc.nasa.gov/data/avhrr/). Other errors related with atmospheric corrections were detected in the PAL database, generating higher values for highest NDVI values (http://daac.gsfc.nasa.gov/CAMPAIGN_DOCS/LA ND_BIO/PAL_coding_errors.pdf). Images were reprocessed. Hence, differences depicted in this work it would not have their origin in this kind of errors. LAC image processing for the analyzed period did not conform to IGBP processing standards (http://edcdaac.usgs.gov/1KM/comp10d.html). The primary difference is in the scaling method applied to byte and integer data. This feature denotes a possible source of divergences between both processing ways. NDVI differences have a non-random spatial variability and the and the higher differences have a defined spatial distribution. The higher differences could be associated to composting processing. Information in low NDVI areas should cautiously analyzed, because associated errors could lead to divergent conclusions in studies of vegetation functioning. Differences found in aisled composites can be attenuated with a correct filtration of the data or with the moving maximum or average of two or more consecutive composites. The use of areas or pixels with constant NDVI (e.g. deserts) could not be an advised practice to solve this kind of differences in NDVI, because there are differences associated to the processing and not to the sensor degradation. The slopes of the adjusted equation and the 1:1 line showed that the differences between data sets are not constant along the NDVI gradient. The use of this equation shows a simple way for the calibration of images coming from the same sensor and with different processing. ACKNOWLEDGEMENTS This work was supported by CONICET, Interamerican Institute for Global Change Studies and Universidad de Buenos Aires. This article is a contribution to the “Plan Estratégico Universidad de Buenos Aires” (15.982/2000 Anexo 33). MFG has a fellowship from University of Buenos Aires. Satellite data were provided by the EOS Data and Information System, DAAC at Goddard Space Flight Center (NASA’s Mission to Planet Earth in cooperation with NOAA). REFERENCES Agbu, P.A and M.E. James, 1994. The NOAA/NASA Pathfinder AVHRR Land Data Set User´s Manual. Goddard DAAC, NASA, Goddard Space Flight Center, Greenbelt. De Fries, R. S., Hansen, M., Townshend, J. R. G., and Sohlberg, R. (1998). Global land cover classification at 8 km spatial resolution: the use of training data derived from Landsat imagery in decision tree clasifiers. Int. J. Remote Sensing 19, 3141-3168. Di Bella, C.M., C.M. Rebella and J.M. Paruelo, 2000; Evapotranspiration estimates using NOAA AVHRR imagery in the Pampa region of Argentina. Int. J. Remote Sensing 21: 791797. Duchemin, B., Guyon, D., and Lagouarde, J. (1999). Potencial and limits of NOAA-AVHRR temporal composite data for phenology and water stress monitoring of temperate forest ecosystems. Int. J. Remote Sensing 20, 895-917. Eidenshink, J.C., and Faundeen, J.L., 1994, The 1km AVHRR global land data set: First stages in implementation. Int. J. Remote Sensing, 15, 3443-3462. James, M. E., and Kalluri, S., 1993, The Pathfinder AVHRR land data set: An improved coarse resolution data set for terrestrial monitoring: Int. J. Remote Sensing, 15, 3347-3364. Loveland; T. R., B. C. Reed; J. F. Brown; D. O. Ohlen; Z. Zhu; L. Yang; J. W. Merchant, 2000. Development of a global land cover characteristics database and IGBP DISCover from 1 km AVHRR data. Int. J. Remote Sensing: 21: 1303 – 1330. NW Pelkey, CJ Stoner, TM Caro, 2000. Vegetation in Tanzania: assessing long term trends and effects of protection using satellite imagery. Biological Conservation: 94, 297-309 Paruelo, J.M. H.E. Epstein, W.K. Lauenroth and I.C. Burke 1997. ANPP estimates from NDVI for the central grassland region of the United States. Ecology 78: 953-958. Tarpley, J.D. S.R. Schneider and R.L. Money 1984. Global vegetation indices from the NOAA-7 meteorological satellite. J.Clim. Appl. Meteorol. 23: 491-494. Tucker, C.J. , S.E. Nicholson, 1999. Variations in the size of the Sahara Desert from 1980 to 1997. Ambio, 28: 587-591. Young, S. S., and Anyamba, A. 1999. Comparison of NOAA/NASA PAL and NOAA GVI Data for vegetation change studies over China. Photogrammetric Engineering & Remote Sensing 65, 679-688