COMPATIBILITY AND INTEGRATION OF NDVI DATA OBTAINED FROM AVHRR/NOAA AND SEVIRI/MSG SENSORS

Transcription

1 COMPATIBILITY AND INTEGRATION OF NDVI DATA OBTAINED FROM AVHRR/NOAA AND SEVIRI/MSG SENSORS Gabriele Poli, Giulia Adembri, Maurizio Tommasini, Monica Gherardelli Department of Electronics and Telecommunication (University of Florence), Via S. Marta 3, Florence, Italy Abstract A statistical analysis on the compatibility of NDVI indices is presented, which is obtained with data collected by AVHRR and SEVIRI (METEOSAT Second Generation MSG) sensors, directly received at the Satellite Receiving Station of Department of Electronics and Telecommunication (DET) - University of Florence, located in Prato (Italy). The goal of this analysis is evaluation of the level of compatibility between the NDVI indices from AVHRR and SEVIRI, so as to extract indications for possibly integrating data from those sensors, exploiting their characteristics (high temporal frequency of available images an image every 15 minutes and the geometric invariance of the Earth-Satellite view for SEVIRI; AVHRR s finest spatial resolution). SEVIRI s and AVHRR s NDVI to be compared were obtained with the same procedure, from data continuously received from January to December 2008 addressing the Tuscany Region, Italy. The processing was carried out using the classical formula (difference and sum ratio of near infrared and red reflectance) and a monthly-based maximum value composite (MVC) technique, so as to limit atmospheric effects. The monthly composite images were then geolocated and registered on Mercator maps with a nearest neighbour interpolation method. Statistical comparison between SEVIRI and AVHRR monthly maps was performed through histograms and scatter plots analysis and through the evaluation of medium values, variances and correlation coefficient. For the considered geographic area and the time period of interest, results show a good concordance between SEVIRI s and AVHRR s NDVI, encouraging the possibility of a combined used of the two sensors. INTRODUCTION NDVI has been a well known and common indicator of vegetation status since 80 s; from the end of 70 s up to today, AVHRR has been widely used for vegetation monitoring purposes; for that reason NOAA data and procedures are a reference, especially for large temporal scale studies (e.g. 10 years or more). The methodology applied to form historical NOAA data series has been recently used for processing data acquired also from new generation sensors (e.g. MODIS, MERIS), thus producing AVHRR compatible NDVI images, meaning they are comparable to NOAA observations (van Leeuwen, 1999) (Huete, 2002) (Chen, 2006) (Günther, 2007). Evaluating the level of compatibility between AVHRR s and other sensors NDVI is useful to investigate the possibility of data fusion and integration. SEVIRI is a new generation sensor that can be efficiently used for NDVI monitoring. The geostationary orbit of MSG and the high temporal resolution of SEVIRI (an image every 15 minutes) are features that increase the probability to have more cloud free images with predefined illumination conditions. This work aims at comparing NDVI monthly maps obtained from SEVIRI and AVHRR, in order to evaluate their compatibility and exploit the key features of the two sensors. The considered geographic zone is the Tuscany Region (Italy) and the period of interest is the year Satellite data used in this work were directly received and processed at the Satellite Receiving Station, located in Prato Campus (lat N, lon E). Since 1997, the Station has been

2 performing direct reception and real time processing of data from NOAA/AVHRR, METEOSAT/MVIRI and, more recently, MSG/SEVIRI. Other important activities of the Receiving Station are the developing of real-time processing software for satellite data (all the software used for this work was specifically developed) and the maintenance and updating of the Station archive, which storages raw and processed data since OBTAINING OF MONTHLY COMPOSITE NDVI MAPS A composite technique MVC (Maximum Value Composite) was used to obtain monthly NDVI maps, in order to minimise the influence of atmospheric effects (clouds, water vapour, aerosols) and the variability of shooting and lighting conditions in evaluation of vegetation index (Holben, 1986) (Gutman, 1989). MVC technique is simple to use: it does not require atmospheric correction of data or complex models based on climatic information, which is not always available with the needed spatial and temporal accuracy. All the factors contributing to satellite measurement contamination (atmospheric effects, sun elevation angle, etc.) have the effect of decreasing the value of NDVI compared to the ideal case. Reliability of the MVC technique is based on the assumption that vegetation in the area of interest does not change significantly during the composition time period. The adopted MVC technique for obtaining monthly SEVIRI and AVHRR NDVI maps on Tuscany Region is described below and illustrated in Figure 1: 1. Choice of composition and observation time intervals. The MVC technique is applied to SEVIRI and AVHRR data on a monthly base; the period of observation is the whole year 2008 (12 months); 2. Selection of AVHRR/NOAA images. Among those received at the Satellite Receiving Station (from NOAA 15, 16, 17, 18), daytime satellite images are selected, that are acquired with a sun elevation angle (El S ) higher than 10 and a look angle (γ) less than 25, namely for at least 90% of the scene (Gutman, 1989). For the selected images, the look angle (γ m ), the sun azimuth and elevation angle (Az S,m, El S,m ) of the centre of the scene are recorded. 3. Selection of SEVIRI/MSG images. Selection of MSG data is based on NOAA data, which were previously selected, i.e. SEVIRI images are selected, whose acquisition time is the closest to the selected AVHRR images time. Since SEVIRI transmits an image every 15 minutes, there will be a difference in acquisition time of at most 7 minutes, between an AVHRR image and the respective SEVIRI image. Also for SEVIRI images, the look angle (which is constant), the sun azimuth and elevation angle of the centre of the scene are recorded (Table 1); Figure 1: Procedure scheme to retrieve NDVI monthly maps (MVC technique) from AVHRR and SEVIRI data. 4. Calculation of NDVI. NDVI is evaluated using reflectance derived from the selected SEVIRI and AVHRR data through equation (1) (NOAA, 1997), where R NIR is the near infrared reflectance (0.6 µm band) and R RED is the red reflectance (0.8 µm band).

3 RNIR RRED NDVI = (1) R + R NIR RED acquisition date [GG.MM.AAAA] AVHRR/NOAA images SEVIRI/MSG images Sun position UTC look NOAA UTC look Diff. Az S,m El S,m acquisition angle γ m satellite acquisition angle γ m AVHRR [deg] [deg] time [hh:mm] [deg] no. time [hh:mm] [deg] [min] : : : : : : : : : : : : : : : : : : : : : : : : : : Table 1: Characteristics of some SEVIRI and AVHRR images selected for composing NDVI map of April Date and time of acquisition, viewing angle averages of AVHRR and SEVIRI sensors (γ m ), minutes of difference than AVHRR acquisition time, average azimuth and elevation of the sun in the scene (Az S,m, El S,m ) are registered for each image. 5. Geolocation and registration with Mercator WGS84 projection. The geolocation of AVHRR data, which has to be accurate as much as possible (Gutman, 1989), is carried out taking into account the orbital model with sub-pixel precision. Data are registered through a nearest neighbour interpolation on a Mercator WGS84 regular grid, whose dimensions are 1156 pixels by 1088 lines (pixel size is km), with a latitude range of 36 00' ' N and a longitude range of 11 45' ' E; 6. Identification of cloudy pixels. A cloud screening procedure is needed, as cloudy pixels have to be discarded. Cloudy pixels are detected, both in AVHRR and SEVIRI maps, using a dynamic threshold test on red and near infrared reflectance (0.6 µm and 0.8 µm) and on brightness temperature (10.8 µm) data. Threshold values are dynamically derived from histograms analysis, as described in (Di Vittorio, 2002); 7. Obtaining of composite NDVI maps. MVC technique is applied to pixels of selected maps; noncloudy pixels, which satisfy both conditions γ<25 and El S >10, are considered for the composition. The condition γ<25 must be checked for every AVHRR pixel, while it is always satisfied for SEVIRI data, since the maximum viewing angle is always less than 8. The final monthly NDVI map is obtained by retaining for each pixel of the Mercator grid the maximum value found for that position among those obtained from single-pass maps processed for a certain month (Figure 1). An example of monthly NDVI composite Mercator WGS84 maps, derived from AVHRR (left) and SEVIRI data (right), is shown in Figure 2. DATA ANALYSIS AND RESULTS The comparison between monthly NDVI indexes obtained from SEVIRI and AVHRR is carried out using a statistical approach. Before performing any statistical computation on NDVI data, those pixels of SEVIRI and AVHRR maps that, in the same location, show a large difference between SEVIRI and AVHRR value, are rejected and not used for the statistics. Such a rejection is performed through the analysis of the scatter plot SEVIRI/AVHRR with geometric moments (Figure 3). In particular, pixels in the maps that show a very different NDVI value between SEVIRI and AVHRR are those generating points in the scatter plot that stand far from its centre. Figure 3 (left) shows, as an example, the scatter plot obtained for January 2008, SEVIRI s NDVI values are on x axis, while AVHRR s are on y axis. The second order moments

4 µ hk of x and y (µ 20, µ 02, µ 11 ) are evaluated with equation (2), where X and Y are the mean values of SEVIRI and AVHRR NDVI. Figure 2: NDVI monthly maps obtained with MVC technique from AVHRR (left) and SEVIRI (right) data (May 2008, Tuscany Region); a Mercator WGS84 projection was used for data registration. Then it is possible to calculate, with equations (3) and (4), axes dimension A, B and slope θ of the major axis of the ellipse characterizing the spatial distribution of data (Teague, 1980). h ( x X ) ( y Y ) µ = (2) hk x, y 2 2 ( µ µ ) + µ ± ( µ µ ) µ A, B = 4 (3) 11 θ = tan µ (4) 00 2 µ 20 µ 02 This ellipse is used to perform the rejection: data standing outside are discarded, inner data are considered for the statistics. Figure 3 (right) shows the map of the rejected points, for January As it could be expected, many of them are placed along the coastline; that is due also to the different ground resolution of the two sensors. For all the months, average percentage of rejected land pixels is about 4-5%. Once the rejection of points has been performed, second order moments (µ 20, µ 02, µ 11 ) are calculated again using only the selected data and for each month of 2008 the following is obtained: - Mean value and standard deviation (Figure 4); - Linear correlation coefficient (Pearson) between SEVIRI and NOAA NDVI (Figure 5); - Histogram of SEVIRI and AVHRR NDVI (Figure 6, top left, top right); - Histogram of SEVIRI AVHRR NDVI difference, with evaluation of difference value at maximum frequency (Figure 6, bottom left); - Parameters (slope and intercept) of major axis of geometric moments ellipse in SEVIRI/AVHRR NDVI scatter plot (Figure 6, bottom right); this line is regarded as the regression line of the scatter plot From monthly mean values, the NDVI annual cycle of year 2008 for Tuscany (Figure 4, left), which can be very useful for gathering information on vegetation. It s immediately possible to note that SEVIRI s and AVHRR s values show the same behaviour, even if SEVIRI s values are always higher than AVHRR s. Such a difference can be related to a slightly less sensitivity of the AVHRR sensor and also to the fact that the pixels of the AVHRR composite image could be acquired with a wider look angle on average. k

5 Figure 3: Rejection of data through analysis of geometric moments; only the data within the ellipse, calculated from second order moments, are taken into account for the statistics (example with data of January 2008). Monthly standard deviation can be related to the dynamic of NDVI observations (Figure 4, right): SEVIRI shows a dynamic range which is globally wider than AVHRR s; in May, vegetation (in its maximum) shows a very uniform distribution. Figure 4: Monthly mean (left) and standard deviation (right) of NDVI from SEVIRI (red) and AVHRR (blue), year In May the maximum value of NDVI is registered, and then June, July and August are the other most vegetative months. Using geometric moments again, the linear correlation coefficient (Pearson) between SEVIRI and NOAA NDVI is obtained with formula (5): µ11 (5) r = µ 20 µ 02 Figure 5 shows the graph of the Pearson coefficient for the year 2008; a good degree of correlation between SEVIRI and AVHRR data can be observed; with a values of r higher than 0.6 for all the months in Figure 6 show the results of the analysis carried out for every month on NDVI data. Histogram of SEVIRI and AVHRR NDVI and of SEVIRI AVHRR NDVI difference can be useful to assess the behaviour of NDVI retrievals.

6 Figure 5: Monthly linear correlation coefficient (Pearson) SEVIRI/AVHRR NDVI, year As for the slope of scatter plot regression lines (i.e. the major axis of the second order moments ellipse), a similar value of about 37 was founded for every month. This is another sign of the good concordance of data, in spite of a different sensitivity. Figure 6: Example of statistical analysis carried out on NDVI data for every month (data of May 2008): histograms of SEVIRI NDVI (top left), AVHRR NDVI (top right), SEVIRI AVHRR NDVI difference (bottom left), scatter plot SEVIRI/AVHRR with regression line (bottom right).

7 CONCLUSIONS AND FUTURE DEVELOPMENTS The aim of this work is to compare the NDVI obtained with SEVIRI and AVHRR data directly received at the Satellite Receiving Station of Prato, applying the same MVC technique. From the analysis of the results we can observe: - SEVIRI sensor results to be more sensitive than AVHRR (higher mean values for the whole year 2008); - SEVIRI sensor shows also a wider dynamic range of NDVI than AVHRR (higher standard deviation values) and this difference increases for the most vegetative months; - There is a pretty good linear correlation between SEVIRI and AVHRR NDVI: r>0.6; - Monthly regression lines with similar slope of about 37. Some differences of behaviour outlined in NDVI values can be attributed to the effect of look angle variations of AVHRR data: the value of NDVI tends to decrease as the look angle increases. Despite MVC technique, there may be residual atmospheric effects that impact differently on the two sensors. Furthermore the statistical analysis method applied can result bounded by the peculiar features of the region and the period of investigation. Combined NDVI monitoring and also reprocessing the available archive data of the previous years will be the goals of future activity. An atmospheric correction improves the reliability of results and can reduce differences between the two sensors; it can also be fruitful to apply the developed procedure to smaller and more uniformly vegetated sub-regional zones. In any case the possibility to go further in the study for data fusion and integration between the two sensors is encouraged by the results obtained in this work. A better ground resolution of AVHRR and the higher availability of SEVIRI data can be exploited through implementation and development of an ad hoc procedures for NDVI data fusion. Another research topic is represented by the assessment of compatibility between SEVIRI NDVI and NDVI from other sensors (e.g. MODIS). ACKNOWLEDGEMENT METEOSAT Second Generation (MSG) data showed in this article originate from the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT). They were received through the license issued by Servizio Meteorologico dell Aeronautica Militare (Ministry of Defence conv. N. 4123/62 of and prot. N.M-DGTEL/4123/2508 of ). REFERENCES Chen, P.-Y., Fedosejevs G., Tiscareño-López M., et Al. (2006) Assessment of MODIS-EVI, MODIS- NDVI and Vegetation-NDVI Composite Data Using Agricultural Measurements: an Example at Corn Fields in Western Mexico. Environmental Monitoring and Assessment, 119/1-3, pp Di Vittorio, A. V., Emery W. J. (2002) An Automated, Dynamic Threshold Cloud-Masking Algorithm for Daytime AVHRR Images Over Land. IEEE Trans. On Geoscience and Remote Sensing, 40-8, pp Günther, K. P., Maier S. W. (2007) AVHRR Compatible Vegetation Index Derived From MERIS Data. International Journal of Remote Sensing, 28/3-4, pp Gutman, G. (1989) On the relationship between monthly mean and maximum-value composite normalized vegetation indices. International Journal of Remote Sensing, 10, pp Holben, B. (1986) Characteristics of maximum-value composite images from temporal AVHRR data. International Journal of Remote Sensing, 7, pp Huete, A. R., Didan K., Miura T., et Al. (2002) Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of Environment, 83, pp

8 van Leeuwen, W. J. D., Huete A. R., Laing T. W. (1999) MODIS Vegetation Index Compositing Approach: Prototype with AVHRR Data. Remote Sensing of Environment, 69, pp National Oceanic and Atmospheric Administration NOAA (1997), NOAA Global Vegetation Index (GVI) User's Guide July 1997 revision. Ed. K. B. Kidwell [ Teague, M. R. (1980) Image analysis via the general theory of moments. Journal of the Optical Society of America B, 70-8, pp