Using ADEOS/POLDER data to reduce angular variability of NOAA/AVHRR reflectances

Remote Sensing of Environment 76 (2001) 399 ± 409

www.elsevier.com/locate/rse

Using ADEOS/POLDER data to reduce angular variability of NOAA/ AVHRR reflectances I. Csiszara,b,*, G. Gutmana,c, P. Romanova,b, M. Leroyd, O. Hautecoeurd b

a NOAA/NESDIS Office of Research and Applications, Washington, DC, USA Cooperative Institute for the Research in the Atmosphere, Fort Collins, CO 80523, USA c NASA Headquarters, Code YS, Washington, DC 20546, USA d CESBIO, UMR CNES-CNR-IPS, Toulouse, France

Received 21 July 2000; accepted 29 December 2000

Abstract Time series of National Oceanic and Atmospheric Administration (NOAA)/Advanced Very High Resolution Radiometer (AVHRR) Global Area Coverage (GAC) data, collected daily over two 50-km targets during March ± June 1997 in Hungary, were corrected for angular effects using coincident multiangle Polarization and Directionality of the Earth's Reflectance (POLDER) Level-2 Land Surface data products. The POLDER data used consisted of narrow-band 0.67- and 0.765-mm reflectances corrected for ozone and water vapor absorption and Rayleigh scattering effects. The AVHRR visible (0.55 ± 0.75 mm) and near-IR (0.68 ± 1.05 mm) data were converted to reflectances, screened for clouds, and corrected for the same atmospheric effects as the POLDER data. Neither POLDER nor AVHRR data were corrected for aerosol effects. POLDER reflectances were used to derive bidirectional reflectance distribution function (BRDF) for each 6-km2 POLDER grid box. The BRDFs were normalized to the near-nadir values at 45° solar zenith angle, resulting in the anisotropic factors, which were derived for each month for each grid box. Thus, seasonal variability of local anisotropy for each POLDER grid box in the target areas was established. The anisotropic factors were then applied to the AVHRR/GAC visible and near-IR reflectances mapped into the POLDER grid. The anisotropy-corrected AVHRR reflectances exhibit less fluctuation than the original uncorrected values, thus, facilitating the interpretation of short-term variability in surface conditions. Application of the POLDER BRDFs to AVHRR data is especially advantageous for processing AVHRR temporal composites because of the scarce angular statistics in the areas of frequent clouds, which hampers derivation of BRDFs from AVHRR data itself. Assuming that the local BRDF does not substantially vary from year to year, this approach could be extended to data from the growing seasons of other years. Ultimately, the multiyear time series could be corrected so that the variability, related to angular effects, inherent to AVHRR time series, is reduced. The current approach suggests a paradigm for a synergistic use of TERRA/EOS Multiangle Imaging SpectroRadiometer (MISR)/MODerate resolution Imaging Spectrometer (MODIS) data stream. D 2001 Elsevier Science Inc. All rights reserved.

1. Introduction The fact that Advanced Very High Resolution Radiometer (AVHRR) time series of land reflectances are fluctuating because of angular variability and the corresponding bidirectional effects has been known for over a decade (e.g., Gutman, 1987; Roujean, Leroy, Deschamps, & Podaire, * Corresponding author. Cooperative Institute for the Research in the Atmosphere, NOAA/NESDIS Office of Research and Applications, E/RAI WWBG 712, 5200 Auth Road, Camp Springs, MD 20746-430, USA. Tel.: +1-301-763-8042 ext. 193; fax: +1-301-763-8108. E-mail address: [email protected] (I. Csiszar).

1992). Noise reduction is important for statistical analysis of time series. In addition to random fluctuation in daily time series, there are biases inherent to long-term weekly, 10-day, or monthly composite datasets (Gutman, 1991) due to preferential sun ± target ± sensor geometry (STSG) induced by certain criteria, such as maximum value compositing. These biases are nonstationary due to satellite orbit drift and can be removed only if the data are normalized to a common STSG. AVHRR data itself can be used to construct bidirectional reflectance distribution functions (BRDFs) only if there is a sufficient number of observations covering the full STSG domain. Attempts have been made to develop empirical or

0034-4257/01/$ ± see front matter D 2001 Elsevier Science Inc. All rights reserved. PII: S 0 0 3 4 - 4 2 5 7 ( 0 1 ) 0 0 1 8 8 - 2

400

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

semiempirical normalization models for AVHRR on a local (Ba, Dedieu, Kerr, Nicholson, & Lecocq, 1997; Cabot & Dedieau, 1997; Cihlar, Chen, Li, Huang, & Pokrant, 1997; Gutman, 1991; Leroy & Roujean, 1994) and global (Zhang, Kalluri, Jaja, Liang, & Townshend, 1998) basis. Difficulties in constructing a BRDF model from AVHRR data arise due to the lack of data for areas with frequent clouds and due to limited STSG coverage. The other disadvantage of the empirical approach is that the statistics are collected over large areas and during an extended time period, hence, large uncertainty in the derived normalization coefficients because of the temporally and spatially variable atmosphere and surface. The Polarization and Directionality of the Earth's Reflectance (POLDER) instrument, which was flown on the Japanese Advanced Earth Observing Satellite (ADEOS) platform during 1996± 1997, provided a unique opportunity to build BRDFs from multiangle synchronous observations. POLDER, a recurrent version of which is scheduled to be flown on ADEOS-2 in late 2001, is described extensively by Deschamps et al. (1994). A surface target is viewed up to 14 times during the satellite overpass, allowing sampling of BRDF at different STSG each successive day with a slightly different directional configuration because of the orbital shift. After a few clear days, therefore, a complete BRDF description can be potentially achieved (Hautecoeur & Leroy, 1998). In areas with frequent clouds, multiangle instruments like POLDER are the only source of data for BRDF construction. Such BRDF, in turn, can be used for normalizing the few AVHRR observations available over the same area. Because of the reduced angular variability, the normalized AVHRR reflectances are more useful in long-term time series analysis. The goal of the present pilot study is to demonstrate the potential of reducing the angular variability observed in AVHRR data by applying POLDER-derived BRDFs. The tests are presented for the 1997 growing season, i.e., when ADEOS/POLDER data were available. Assuming that the local BRDF does not substantially vary from year to year, unless anomalous conditions occur, such as drought or flood, this approach could be extended to the data for growing seasons of other years. Ultimately, the multiyear AVHRR time series could be corrected so that the variability inherent to AVHRR time series is reduced. Results of the current pilot study could be utilized as a paradigm for a synergistic use of TERRA/EOS Multiangle Imaging SpectroRadiometer (MISR) and MODerate resolution Imaging Spectrometer (MODIS) data (Wanner et al., 1997). 2. Data The current study used daily time series of POLDER Level-2 Land Surface product and of AVHRR Global Area Coverage (GAC) observations. We used GAC data from a

special daily dataset that has been collected at the National Oceanic and Atmospheric Administration, National Environmental Satellite Data and Information Service (NOAA/ NESDIS) over two 50  50-km target areas in Hungary: (1) near Szarvas centered at 46.8°N, 20.5°E, which is mainly cropland; and (2) near Keszthely centered at 46.8°N, 17.2°E, which is a mixed area containing forests, gardens, and marshland. Only data from the growing season (March ± June) were considered in this study. The two target areas represent midlatitude vegetated surfaces, with a considerable number of cloudy observations. They thus serve as typical examples of areas where the proper preprocessing of the data, including the removal of data variability due to angular effects, is essential to monitor surface conditions. POLDER Level-2 Land Surface data represent narrowband (0.02 ± 0.04 mm bandwidth) reflectances at 0.443, 0.67, 0.765, and 0.865 mm, mapped onto a 1/18° grid. The 0.765-mm reflectance is from a standard radiance measurement, whereas at 0.443, 0.67, and 0.865 mm, polarized intensities are taken at three polarization angles differing by 60° between successive polarizers. The procedure to derive radiances from the polarized intensities has been shown to have insignificant errors, and thus, for our purposes, there is no difference between manipulating radiances (and reflectances) that come from polarized measurements or those originating from standard measurements. The data were calibrated, cloud-screened, corrected for Rayleigh scattering and absorption by ozone and water vapor at the preprocessing stage (Leroy et al., 1997). No correction for aerosol, however, was done at this time. For this study, Channels 2 (0.67 mm) and 3 (0.765 mm) were selected due to their proximity to the peaks of the much broader NOAA-14/AVHRR visible and near-IR spectral response functions at 0.68 and 0.78 mm, respectively (Kidwell, 1998). Preprocessing of AVHRR Channels 1 (visible) and 2 (near-IR) data included conversion of raw counts to albedo units with updated postlaunch NOAA-14 calibration (Rao and Chen, 1999), cloud screening, and atmospheric correction. Cloud screening was based on the technique described in Gutman (1991), i.e., application of gross thresholds and fine screening based on spatial variability and restricting visible reflectance to be within 1 S.D. of the mean for each viewing angle bin for each month. Rayleigh, ozone, and water vapor corrections were conducted with the SMAC code (Rahman & Dedieu, 1994). The total water vapor amount was derived from radiosonde data available at the meteorological stations near the target areas. No aerosol correction was made for consistency with the POLDER product. AVHRR data were mapped onto the POLDER grid. Fig. 1 shows the STSG domains of clear POLDER and AVHRR observations for the 4 months over the Szarvas target area (conditions over the Keszthely target area are very similar). As solar zenith angle (qs) has less variability

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

401

Fig. 1. (a) Monthly distribution of satellite viewing angle (qv) and relative azimuth angle (F) of clear POLDER observations over the Szarvas target area at the corresponding approximate mean of solar zenith angle (qs). Negative values of qv correspond to backscatter conditions. (b) Same as (a) for AVHRR observations.

within a given month, data are shown in the satellite viewing angle (qv) ± relative azimuth angle (F) plane. Negative values of qv correspond to the backscatter conditions. POLDER data (Fig. 1a) form distinct group of dots for each day, whereas AVHRR observations (Fig. 1b) are represented by one dot for a day. It can be seen that within a few days POLDER observations cover practically the entire STSG domain. This rich angular sampling cannot be reproduced even by a long time series of AVHRR observations, due to a limited range of relative azimuth angles. Fig. 2 illustrates the distribution of clear visible and nearIR POLDER and AVHRR data with viewing angle. POLDER data are shown only for the Szarvas target area (Fig. 2a and b). The comparison of POLDER and AVHRR data for Szarvas reveals the large differences in data amount. The wide range of POLDER reflectances corresponding to distinct values of qv is a result of their angular (mostly relative azimuth), spatial, and temporal variations over the target areas within the given month. On the other hand, the range of AVHRR data within a narrow range of qv, usually representing only 1 day, characterizes only spatial heterogeneity. Temporal changes of reflectance can be caused by both surface changes and by residual atmospheric effects. The stronger spatial variability and the general increase of near-IR reflectances during the growing season show the sensitivity of these measurements to vegetation development. The monthly variation of near-IR AVHRR reflectances with qv is thus the net result of angular effects and vegetation growth. It introduces further uncertainties in the BRDF derivation if only AVHRR data are used. AVHRR

data also show that, in some cases, even a month worth of data may not contain clear observations covering the whole viewing angle domain, which would be necessary for the construction of a statistically stable BRDF. Note also that in the forward scattering direction, the angular distribution function of AVHRR reflectances is much more variable, with many scattered data points, which suggests that residual cloud contamination and/or specular reflection from water may still be present in the data. A weak ``hot spot'' signal (increase of reflectance at near-coincident solar illumination and viewing angles) is observable in May and June in the visible, whereas in the near-IR, it is obscured by the faster temporal changes. (The ``hot spot'' signal could be seen in the POLDER data only with data near the principal plane.) Note also that in April, when there are only a few AVHRR observations, POLDER data are also missing for near-nadir viewing angles. 3. Method 3.1. Parameterization of BRDF The derivation of BRDFs in this study is based on the method used by Leroy et al. (1997), Leroy and Roujean (1994), and Wu, Li, and Cihlar (1995). It relies on the representation of a local BRDF by a Kernel-driven model by Roujean, Leroy, and Deschamps (1992). The BRDF is characterized by two basic physical functions parameterizing two main effects: (1) geometric (shadowing) effect and (2) volume scattering. The directional spectral reflectance in

402

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

Channel i at solar zenith angle Qs, viewing zenith angle Qv, and relative azimuth angle F, Ri(Qs,Qv,F), is described by the formula (Eq. (1)): Ri …Qs ; Qv ; F† ˆ k0i ‡ k1i f1 …Qs ; Qv ; F† ‡ k2i f2 …Qs ; Qv ; F†

…1†

where k0i = Ri(0°,0°,F) is the nadir reflectance at overhead sun (note that for nadir view and/or illumination, the notation of the dependence on relative azimuth angle is only

a formalism). Analytical description of the Kernel functions f1 and f2 was given in Roujean et al. (1992). k0i, k1i, and k2i coefficients are found from regression of data on f1 and f2.The BRDF model Vi(Qs,Qv,F) is defined as (Eq. (2)): Vi …Qs ; Qv ; F† ˆ Ri …Qs ; Qv ; F†=Ri …0; 0; F† ˆ 1 ‡ a1i f1 …Qs ; Qv ; F† ‡ a2i f2 …Qs ; Qv ; F†

…2†

where a1i = k1i/k0i and a2i = k2i/k0i. Further, we define an anisotropic factor hi(Qs,Qv,F) that is used for normalization

Fig. 2. (a) Distribution of clear visible POLDER data with viewing angle (qv) over the Szarvas target area. (b) Distribution of clear near-IR POLDER data with viewing angle (qv) over the Szarvas target area. (c) Distribution of clear visible AVHRR data with viewing angle (qv) over the Szarvas target area. (d) Distribution of clear near-IR AVHRR data with viewing angle (qv) over the Szarvas target area. (e) Same as (c) for the Keszthely area. (f) Same as (c) for the Keszthely area.

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

403

Fig. 2. (continued).

of an AVHRR measurement to a reference STSG at nadir (Qv = 0°) with Qs = 45° (Eq. (3)): hi …Qs ; Qv ; F† ˆ Vi …Qs ; Qv ; F†=Vi …45°; 0°; F†

…3†

3.2. Quality control of POLDER data Fig. 2 shows some outlying POLDER reflectances, most clearly seen in June. This indicates that some residual atmospheric effects (residual clouds and/or unusually high aerosol content) are still present in the input data. The outliers cause distortions in the BRDF and thus outlying k0i, k1i, and k2i values for those locations and dates. To control the quality of the data input to BRDF construction, k0i, k1i, and k2i coefficients were derived for each 1/18° grid box within the two targets for each day available in the cloud-free Level-2 POLDER data. We applied further screening of the data by restricting k0i, k1i, and k2i to be within 1 S.D. from the means of the population over 1 month for each 50-km target. This screening method implies assumptions that: (1) the data are mostly cloud-free; (2) the

spatial variability of the BRDF effects can be neglected in comparison to the residual atmospheric effects; and (3) the BRDF model used works reasonably well with data from any geometric sample. The latter assumption is probably not very solid, but this is the only solution that we could arrive to under the present circumstances, i.e., when the atmospheric effects are still significant in the POLDER Level-2 surface data products. The above screening was applied two times, yielding the values of k0i, k1i, and k2i at the second iteration very close to those obtained at the first step. Fig. 3 shows the effect of the screening for June. The regression for the original data (left panel) produces some incorrect estimates, seen here as outlying data points. The screening removed the outliers, along with some data with higher reflectances, which can potentially be cloud-contaminated pixels. The range of visible POLDER reflectances thus became much similar to that of AVHRR reflectances in that month (Fig. 2c). The latter, including data close to the principal plane in the backscatter direction, is not expected to be surpassed by the range of POLDER data at any other STSG. Scrutiny of the data also showed that the screening mostly removed

Fig. 3. Predicted vs. observed visible POLDER reflectances before (left) and after (right) the additional screening of the input dataset.

404

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

data corresponding to 1 particular day (June 26). Note that some cloud-free data may also have been removed by the conservative screening procedure. However, the number of remaining POLDER observations (e.g., > 800 for June for the Szarvas target area) and the wide range of STSG (very similar to that shown in Fig. 1a) still provide sufficient data sampling for deriving BRDFs. 3.3. Derivation of BRDFs from POLDER and AVHRR data Coefficients corresponding to POLDER data remaining after the quality control stage were averaged for each month within each grid box and used in deriving the final anisotropic factors. For comparison, we also attempted to derive monthly anisotropic factors for each grid box from AVHRR data. Statistical analysis of the data, however, revealed that for the vast majority of the grid boxes, the number of cloud-free AVHRR observations was below five, with the maximum numbers being 13 and 8 for the Szarvas and Keszthely target areas, respectively. The poor statistics of the data thus did not allow to perform the regression for each grid box separately. Instead, we derived monthly BRDFs from all cloud-free AVHRR observations within the given month for each 50  50km target area. 4. Results Fig. 4 shows the seasonal variability of the coefficients k0i, k1i, and k2i during the growing season for both POLDER and AVHRR. The POLDER graph shows spatial means and their standard deviations, whereas the AVHRR

graph shows the single values derived for each month, hence, no standard deviations. The physical interpretation of the seasonal variability in the coefficients follows from their parameterization. The normalized reflectance, k0, for visible POLDER data shows a decrease from March values of 0.07 (Szarvas) and 0.09 (Keszthely), to 0.05 ± 0.06 in June. The normalized near-IR POLDER reflectance shows a steady increase from 0.15 (Keszthely) and 0.17 (Szarvas) to 0.25 ±0.27, with systematically higher values during the growing season for the forested Keszthely area. The first loading, k1, characterizing the geometric effects, shows a less regular behavior than the two other coefficients. Its value is close to 0, for several months taking unphysical negative values, similarly to findings by Roujean et al. (1992) and Wu et al. (1995) for cropland and forest. Statistical analysis, however, showed that the deviation of the mean k1 coefficients from 0 is not significant. In a physical sense, the contribution of the first Kernel function is negligible over the two target areas of this study. The second loading, k2, characterizing the volume scattering, which could possibly be interpreted as an increase in leaf area index, exhibits a decrease from March to April ± June for the visible, and a steady increase starting in April for the near-IR POLDER data. The systematically higher values of near-IR k2 values for the Keszthely area as compared to Szarvas is explained by more abundant vegetation there. The general behavior of the AVHRR coefficients is similar to that of POLDER's, except for spurious values in April, which is explained by the extremely poor statistics because of the lack of cloud-free days (see Fig. 2c and d). The differences between the coefficients for the two target areas are significantly larger, caused at

Fig. 4. Temporal development of the coefficients k0i, k1i, and k2i during the growing season derived from POLDER (left) and AVHRR (right) data.

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

least partly by distortions from the poor angular representation of AVHRR data. Fig. 5 shows a comparison of anisotropic factors as a function of the viewing zenith angle for solar zenith and relative azimuth angles typical to AVHRR measurements, derived from POLDER and AVHRR for each month. The POLDER-derived BRDFs were calculated from monthly means of the k0i, k1i, and k2i coefficients for this illustration. They are very similar for the two target areas. They are also very close to each other for May and June, i.e., when

405

vegetation is well developed. The April BRDFs differ only slightly from those of May ± June, but the ones for March are significantly different because of the difference between the anisotropic properties of bare soil and vegetation. The AVHRR BRDF's behavior is more irregular than that of POLDER's. The stability of the AVHRR-derived BRDFs is directly related to the quality and amount of input data (Fig. 2c ± f). The poor statistical sample used in the regression yields stronger (Szarvas: all months in the visible, and April in the near-IR) or weaker (Keszthely: May ± June in the

Fig. 5. (a) Visible anisotropic factors vs. the satellite viewing angle (qv) for solar zenith (qs) and relative azimuth (F) angles typical to AVHRR measurements, derived from POLDER and AVHRR data for Szarvas. The F values denote the relative azimuth angles for backscatter and forward scatter for which the anisotropic factors are shown. (b) Near-IR anisotropic factors vs. the satellite viewing angle (qv) for solar zenith (qs) and relative azimuth (F) angles typical to AVHRR measurements, derived from POLDER and AVHRR data for Szarvas. The F values denote the relative azimuth angles for backscatter and forward scatter for which the anisotropic factors are shown. (c) Same as (a) for Keszthely. (d) Same as (b) for Keszthely.

406

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

visible, April in the near-IR) angular dependences than the ones derived from POLDER. The AVHRR BRDFs for several months, however, are very close to the corresponding POLDER-derived functions (e.g., near-IR in June for both areas). This is not surprising if one considers the relatively large data amount in June and the well-defined, uniform angular distribution of reflectances. The results of application of POLDER- and AVHRRderived anisotropic factors to AVHRR time series are shown in Fig. 6. For this illustration, the observed and normalized AVHRR reflectances were grouped into backscatter (Qv < 20°), near-nadir ( 20° < Qv < 20°), and forward scatter (20° < Qv) bins and spatially averaged for each bin for each day and for each target area (as the target areas were relatively small, for most days, data fell into only one angular bin). At a first glance, it is evident that the corrected reflectances remain within the envelope defined by the forward scatter (triangles) and backscatter (circles) observations, and, indeed, there is less fluctuation in the corrected time series. The POLDER normalization was more efficient for the near-IR than for the visible. This is because the inherent atmospheric noise in the visible is still comparable with the BRDF effects. That is, even after the additional screening procedures, some of the contamination remained in the data; hence, the BRDF corrections were not efficient in those cases. For example, three dates in Szarvas (Julian days 73, 93, and 120) with uncorrected data points in the visible are especially noticeable. The observed reflectances, however, appear to be too high for the forward scatter direction, which suggests residual atmospheric effects. Therefore, we examined AVHRR imagery obtained from the NOAA Satellite Active Archive (www.saa.noaa.gov).

On day 73 (March 14), GAC imagery shows cloud-free, but distinctively hazy conditions over the Szarvas target area. On day 93 (April 3), the Szarvas target area is again predominantly clear on the 1-km Local Area Coverage (LAC) imagery. However, several fire hot spots appear in the area, with a feature appearing as a smoke plume originating from an extensive hot spot including three 1km pixels. The suggested aerosol effect is consistent with the fact that Channel 1 reflectance is effected more than Channel 2. On day 120 (April 30), the area is mostly cloudy, with open patches around Szarvas. While the presence of atmospheric haze could not be confirmed, we observed small-scale cumulus cloudiness, which remained undetected by the cloud screening of GAC data described in Section 2. Note that this cloud screening technique did not account for cloud shadow effects either, which can be a significant problem in predominantly cloudy areas. The fact that data discussed in the above examples were undercorrected by POLDER confirms the stability of the POLDER-derived BRDFs and their potential to screen out AVHRR data that are contaminated by residual atmospheric effects. The AVHRR-derived BRDFs, however, now applied to the same statistically dependent AVHRR data sample, normalized these contaminated data (Fig. 6b), using much higher values of anisotropic factors (  1.9 vs.  1.05 from Fig. 5a). Apart from the several examples above, the correction by the AVHRR-derived BRDFs was also successful, occasionally, even superior to the POLDER-derived corrections (e.g., May in Keszhely in the near-IR). However, the much richer angular sampling of POLDER suggests that, in many cases, the AVHRR-derived BRDFs, especially when applied to an

Fig. 6. (a) Time series of observed and corrected reflectances using POLDER-derived anisotropy factors over the two target areas. Reflectances were averaged daily for each satellite viewing angle bin. (b) Same as (a) for AVHRR-derived anisotropy factors.

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

407

Fig. 7. Comparison of time series of hemispheric albedos derived from AVHRR and POLDER measurements.

independent sample of AVHRR data, would produce wrong angular corrections and incorrect values of hemispheric albedos after the integration over the entire angular domain. To demonstrate this, we derived time series of hemispheric albedos from AVHRR and POLDER BRDFs for the solar zenith angles of the AVHRR observations using the parameterization by Roujean et al. (1992). The comparison of the time series (Fig. 7) shows that for the months of apparently unrealistic BRDFs from AVHRR (most noticeably in April in Szarvas in the visible, and in Keszthely in the near-IR), the AVHRR-derived hemispheric albedos

significantly deviate from the smoother time series of POLDER-derived albedos. Note also that the decrease of Qs during spring causes a decreasing trend of visible albedos, not seen in the time series of reflectances (Fig. 6a and b), where the data were normalized to a reference Qs. In the near-IR, the sun angle effect is obscured by the increase of reflectances due to vegetation growth. The residual noise caused by atmospheric effects can be removed also by statistical means. Statistical filtering is useful, but should be applied after the physical corrections have been applied. Fig. 8 shows the results of the application of a three-point median filter to time series of original and corrected reflectances (Fig. 6). The curves of the corrected reflectances are much smoother than those obtained by smoothing the original observations, because the majority of the data points were indeed corrected. However, significant differences between the smoothed curves of AVHRR- and POLDER-corrected reflectances exist, especially in the near-IR. This suggests that the smoothness of the restituted temporal profiles is not the only criterion to assess the quality of the corrected data. Further analysis, including comparison with in situ data wherever possible, is needed to identify any consistent biases of the applied angular correction procedure. 5. Conclusions

Fig. 8. Smoothed time series of observed and POLDER (left)- and AVHRR (right)-corrected reflectances by a median filter.

This pilot study demonstrates the potential of deriving local anisotropic factors from POLDER observations and applying them to AVHRR time series. The results show that the fluctuation in the AVHRR time series was decreased, which is important for statistical studies. The results are

408

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409

especially effective in the near-IR band. The variability in the visible data is reduced less significantly because of the residual atmospheric effects, which are comparable in magnitude to BRDF effects. An important conclusion from our experience of working with the current POLDER Level-2 data product is that much additional screening is required to eliminate residual atmospheric effects. This is dictated by the sensitivity of the BRDF model used here to uncertainties in the data input for regression analysis. It is foreseen to implement an additional filtering process in a forthcoming version of the processing of POLDER data, to be applied to POLDER/ADEOS-2 and retrospectively to POLDER/ ADEOS-1 data. On the other hand, a statistically stable BRDF model from POLDER can be used to identify and eliminate contaminated data in the AVHRR time series. The BRDF signal can also be used in aerosol retrieval algorithms over land. Derivation of monthly BRDFs from AVHRR data only on a < 10-km resolution is not possible within the climatic conditions of the areas studied. Collection of longer (i.e., more than a month) AVHRR or MODIS time series data to yield a larger number of observations covering a somewhat wider STSG domain would be unreliable for periods and areas of relatively fast surface changes, especially in the near-IR. Sufficient number of AVHRR observations necessary for the construction of realistic monthly BRDFs can be achieved only by two ways. One is by increasing the area for which one single BRDF is derived, up to the scale of the 50  50-km targets according to our experience. This procedure, however, eliminates the signal from the spatial variability of the data. Another possibility is the collection of data from different years from a smaller area (Cihlar et al., 1997). This approach probably produces more representative local BRDFs, but makes the detection of small interannual changes much more difficult. The poor AVHRR data sample, in addition, may still include cloud- and water-contaminated data. The regressions used for the derivation of the kij loadings produced significantly distorted and unstable BRDFs to satisfy the few existing data points. The time series of AVHRR-derived BRDF corrections demonstrated that cloud- and aerosolcontaminated data were ``corrected.'' In the POLDER dataset, which includes far more data points, any possible contaminated or otherwise unusable data have very little effect on the statistically stable regression. Therefore, multiangle sensors, such as POLDER or MISR, produce statistically much more stable BRDFs than single-angle observations, such as AVHRR or MODIS. This is especially important if hemispheric albedos are derived. In frequently cloudy areas, POLDER- or MISR-derived BRDFs will be indispensable and will provide the only tool for noise reduction in long-term AVHRR time series. Future work will include testing application of the BRDFs derived for the 1997 growing season to other years. Similar studies could and should be made over other land

surface types. The current approach could also be tested with the forthcoming data stream from TERRA/EOS MISR and MODIS for developing a system for a synergistic use of their data in land cover change studies. MISR BRDFs could be derived during the first year and then applied to MODIS observations, retrospectively, as well as in real time during the second year. Acknowledgments We thank Dr. J.-L. Roujean (Meteo-France) for his valuable comments on the manuscript. Thanks also go to J. Kerenyi for preprocessing AVHRR data during her stay at NOAA/NESDIS Office of Research and Applications, on leave from the Hungarian Meteorological Service. References Ba, M. B., Dedieu, G., Kerr, Y. H., Nicholson, S. E., & Lecocq, J. (1997). Reduction of bidirectional effects in NOAA ± AVHRR data acquired during the HAPEX-Sahel experiment. Journal of Hydrology, 188 ± 189, 725 ± 748. Cabot, F., & Dedieu, G. (1997). Surface albedo from space: coupling bidirectional models and remotely sensed measurements. Journal of Geophysical Research, 102, 19645 ± 19663. Cihlar, J., Chen, J., Li, Z., Huang, F., & Pokrant, H. (1997). Can interannual land surface signal be discerned in composite AVHRR data? Journal of Geophysical Research, 103, 23163 ± 23172. Deschamps, P. Y., BreÂon, F. M., Leroy, M., Podaire, A., Bricaud, A., Buriez, J. C., & SeÁze, G. (1994). The POLDER mission: instrument characteristics and scientific objectives. IEEE Transactions on Geoscience and Remote Sensing, 32, 598 ± 615. Gutman, G. (1987). The derivation of vegetation indices from AVHRR data. International Journal of Remote Sensing, 8, 1235 ± 1243. Gutman, G. (1991). Vegetation indices from AVHRR: An update and future prospects. Remote Sensing of Environment, 35, 121 ± 136. Hautecoeur, O., & Leroy, M. (1998). Surface bidirectional reflectance distribution function observed at global scale by POLDER/ADEOS. Geophysical Research Letters, 25, 4197 ± 4200. Kidwell, K. B. (Ed.). (1998). NOAA polar orbiter data users guide. US Department of Commerce, National Oceanic and Atmospheric Administration, National Environmental Satellite Data and Information Service, (pp. 1-70 ± 1-71). Suitland, MD: National Climatic Data Center. Leroy, M., DeuzeÂ, J. L., BreÂon, F. M., Hautecoeur, O., Herman, M., Buriez, J. C., TanreÂ, D., BouffieÁs, S., Chazette, P., & Roujean, J. L. (1997). Retrieval of atmospheric properties and surface bidirectional reflectances over land from POLDER/ADEOS. Journal of Geophysical Research, 102, 17023 ± 17037. Leroy, M., & Roujean, J. L. (1994). Sun and view angle corrections on reflectances derived from NOAA/AVHRR data. IEEE Transactions on Geoscience and Remote Sensing, 32, 684 ± 696. Li, Z., Cihlar, J., Zheng, X., Moreau, L., & Ly, H. (1996). The bidirectional effects of AVHRR measurements over boreal regions. IEEE Transactions on Geoscience and Remote Sensing, 34, 1308 ± 1322. Rahman, H., & Dedieu, G. (1994). A Simplified Method for the Atmospheric Corrections (SMAC) of satellite measurements in the solar spectrum. International Journal of Remote Sensing, 15, 123 ± 143. Rao, C. R. N., & Chen, J. (1999). Revised post-launch calibration of the visible and near-infrared channels of the Advanced Very High Resolution Radiometer (AVHRR) on the NOAA-14 spacecraft. International Journal of Remote Sensing, 20, 3485 ± 3491.

I. Csiszar et al. / Remote Sensing of Environment 76 (2001) 399±409 Roujean, J. L., Leroy, M., & Deschamps, P. Y. (1992). A bidirectional reflectance model of the Earth's surface for the correction of remote sensing data. Journal of Geophysical Research, 97, 20455 ± 20468. Roujean, J. L., Leroy, M., Deschamps, P. Y., & Podaire, A. (1992). Evidence of surface reflectance bidirectional effects from NOAA/AVHRR multitemporal data set. International Journal of Remote Sensing, 13, 685 ± 698. Wanner, W., Strahler, A. H., Hu, B., Lewis, P., Muller, J.-P., Li, X., Barker Schaaf, C. L., & Barnsley, M. J. (1997). Global retrieval of

409

bidirectional reflectance and albedo over land from EOS MODIS and MISR data: theory and algorithm. Journal of Geophysical Research, 102, 17143 ± 17162. Wu, A., Li, Z., & Cihlar, J. (1995). Effects of land cover type and greenness on AVHRR bidirectional reflectances: analysis and removal. Journal of Geophysical Research, 100, 9179 ± 9192. Zhang, Z., Kalluri, S. N. V., JaJa, J., Liang, S., & Townshend, J. R. G. (1998). Models and high-performance algorithms for global BRDF retrieval. IEEE Computer Science and Engineering, 5 (4), 16 ± 29.