Elsevier

Remote Sensing of Environment

Volume 169, November 2015, Pages 390-403
Remote Sensing of Environment

Evaluation of the Landsat-5 TM and Landsat-7 ETM + surface reflectance products

https://doi.org/10.1016/j.rse.2015.08.030Get rights and content

Highlights

  • We evaluate LEDAPS-derived Landsat-5 TM and −7 ETM + surface reflectance (SR).

  • We evaluate the theoretical performances of LEDAPS over 489 AERONET sites.

  • We build a methodology to compare MODIS and Landsat SR with BRDF and spectral adjustments.

  • The Landsat SR uncertainty is below the specification (0.05ρ + 0.005) except for blue band.

  • The Landsat SR uncertainty is stable through time.

Abstract

Maintaining consistent datasets of Surface Reflectance (SR) is an important challenge to ensure long-term quality of Climate Data Records. The Landsat 5 and 7 archives offer a unique data source to monitor globally the land surface at high spatial resolution. The Landsat-5 TM and Landsat-7 ETM + SR products, derived from the on-demand processing Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS), require periodic evaluation to check the data consistency. Two evaluation approaches are presented in this paper. The first approach used the Aerosol Robotic Network (AERONET) data set for the period 2000 to 2013 over 489 sites with 3600 Landsat-5 TM and Landsat-7 ETM + scenes selected. For each scene, 10 × 10 km subsets of LEDAPS-derived Landsat SR and AERONET-derived SR are compared. The latter are computed using Landsat top of atmosphere reflectance, AERONET measurements of atmospheric parameters, and the 6S radiative transfer model. Second, we introduce a methodology to cross-compare Landsat data and MODIS data acquired on the same day. The analysis is based on 4000 random Landsat scenes globally distributed from 2000 to 2013. This method includes: (i) a surface anisotropy adjustment, based on the VJB Bidirectional Reflectance Distribution Function (BRDF) method, to adjust Terra and Aqua MODIS data to Landsat 5 and 7 sun-view geometry, (ii) a spectral adjustment based on an artificial neural network trained with the PROSAIL vegetation radiative transfer model, to adjust MODIS data to TM and ETM + spectral responses.

The overall results of both approaches show a good match in over 80% of the scenes, i.e. the TM and ETM + SR uncertainty remained within the SR specification, defined as 0.05 × SR + 0.005. The worst results are found in the blue band used in LEDAPS to adjust the Aerosol Optical Thickness (AOT). The MODIS-Landsat SR cross-comparison confirms the utility of a BRDF adjustment method to decrease the scattering between Landsat sensors and MODIS sensors (Terra and Aqua). The spectral adjustment removes part of the biases related to spectral response differences. Global analysis is used to identify AOT retrieval issues over specific scenes, mostly over bright surfaces. From 2000 to 2013, no significant temporal variation of the performance is detected, which enhanced the consistency of LEDAPS-derived surface reflectance data set.

Introduction

Since 2009, the Landsat data archive has been opened at no cost to the scientific community (Woodcock et al., 2008). It offers a unique data source to globally monitor the land surface at a high spatial resolution. With more than 42 years of data, Landsat archive is a valuable data set for analyzing long-term land surface dynamics (Roy et al., 2014b). In 2012, the USGS (United States Geological Survey) Earth Resources Observation and Science Center (EROS) released a Climate Data Record (CDR) corresponding to a global land surface reflectance (SR) coverage from Landsat archived images that met quality and cloud cover specifications (http://landsat.usgs.gov/CDR_LSR.php). USGS processed the Global Land Survey (GLS, Gutman, G., et al., 2008, Tucker, C.J., et al., 2004) for year 2000, 2005 and 2010 using the atmospheric correction chain LEDAPS (Landsat Ecosystem Disturbance Adaptive Processing System, Masek et al., 2006). In 2010, USGS-EROS released on-demand LEDAPS processing for generating SR products for the entire Landsat-5 and Landsat-7 archives (http://espa.cr.usgs.gov).

The current version of LEDAPS was developed to generate automatic SR products from Landsat-5 Thematic Mapper (TM) and Landsat-7 Enhanced Thematic Mapper Plus (ETM +) calibrated Top Of Atmosphere (TOA) reflectance. The algorithm relies on: (i) the Second Simulation of a Satellite Signal in the Solar Spectrum (6S) radiative transfer model (Vermote, Tanre, Deuzé, Herman, & Morcrette, 1997); (ii) auxiliary data sources for water vapor, air pressure and air temperature (from the National Centers for Environmental Prediction - NCEP), ozone (from the NASA Earth Probe Total Ozone Mapping Spectrometer), and topography; (iii) internal aerosol optical thickness (AOT) retrieval. The latter is retrieved using Kaufman et al. (1997) Dense Dark Vegetation (DDV) approach for each individual Landsat scene, combined with a fixed continental aerosol model. DDV pixels are defined by a threshold on spectral TM and ETM + blue and red bands (see Fig. 1 for band names and wavelengths). AOT at 550 nm are retrieved over these DDV pixels assuming a constant relationship between these two spectral bands. AOT are finally mapped through the image using bilinear or nearest-neighbor interpolation, depending on the distance. A 0.06 AOT value is attributed to pixels further than 7 km of any DDV pixels.

The development of a long-term surface reflectance CDR requires the development and implementation of methods to verify the quality of the product (Quality Assurance — QA) and its accuracy (validation). The QA is undertaken in this work by cross comparison with MODIS SR product that have been adequately characterized by rigorous QA (Roy et al., 2002) and validation (Vermote & Kotchenova, 2008). The validation of the SR in this work follows MODIS methodology which rely on using the Aerosol Robotic Network (AERONET) sites (Holben et al., 1998) and rigorous radiative transfer (Kotchenova, S.Y. and Vermote, E.F., 2007, Kotchenova, S.Y., et al., 2006, Vermote, E.F., et al., 1997). QA and validation of LEDAPS surface reflectance has been discussed by three studies prior to this current one: Ju, Roy, Vermote, Masek, and Kovalskyy (2012); Feng et al. (2013) and Maiersperger et al. (2013). They addressed this scientific question using two QA approaches (assessment of the LEDAPS modeling approach and cross-comparison with MODIS SR data) and two validation approaches (validation over the AERONET sites and vicarious field spectrometer comparisons).

Since 1990, the international AERONET has provided a long-term, continuous and readily accessible public domain database of aerosol optical, microphysical and radiative properties (Holben et al., 1998). The network offers a unique way to validate SR products by using accurate atmosphere characterization measurements (Vermote & Kotchenova, 2008). Maiersperger et al. (2013) used 95 AERONET sites located in North America to compare LEDAPS aerosol optical thickness (AOT) retrieval with AERONET AOT. The Landsat data set corresponded to 3514 scenes from 1993 to 2011. They found that 57% of the sites have a median absolute AOT deviation below 0.1. They observed an overall overestimation of AOT on sparse vegetation, arid lands, or near water where DDV retrieval was problematic (overall.). Ju et al. (2012) used 26 US AERONET sites to compare SR from 82 Landsat ETM + images acquired in 2007–2008. They calculated a relative uncertainty of 11.8% for the blue band, of the order of 6% for green and red bands and less than 5% for infrared bands. More specifically for bands blue to near-infrared (NIR), they observed an underestimation of low values and overestimation of high values.

Ju et al. (2012) have evaluated the accuracy of LEDAPS atmospheric correction by comparing the performance LEDAPS SR with the performances of SR retrieved from a MODIS-based atmospheric correction method applied on the same Landsat TOA (Top of Atmosphere) data. Both methods are based on the 6S radiative transfer code but differ in the AOT retrieval approaches. The results indicated that the MODIS-based method had overall higher accuracy than the LEDAPS method for all the ETM + bands except for the green band, where the results for the two methods were comparable, and the blue band, where both the LEDAPS and MODIS-based atmospheric correction methods performed less reliably.

As of 2014, the Maiersperger et al. (2013) study is the only LEDAPS-derived SR vicarious validation study using field spectrometer measurements. LEDAPS values were reasonably corroborated over test sites surrounded by DDV, with an overall relative difference of 3% for all spectral bands except 1600 nm band (4%). However, this time-consuming approach was carried out with a very limited number of Landsat scenes (16) distributed over three sites.

Another QA approach, addressed by Feng et al. (2013) and Maiersperger et al. (2013), is the cross-comparison of SR products derived from independent sensors. The two studies, which focus on cross-comparison with MODIS SR products, assume a linearity during the aggregation process from the 30 m Landsat pixel to the coarser MODIS pixel footprint (> 500 m). Maiersperger et al. (2013) compared Landsat-7 ETM + and MODIS Terra NDVI over 4 sites located in USA resulting in a total of 48 images. This comparison did not require any adjustment of the surface anisotropy effects since both platforms are in the same orbit with a 30-minute lag (i.e., almost no difference in terms of sun-view geometry). Based on more than 18 thousand Landsat scenes derived from the 2000 and 2005 Global Land Surveys, Feng et al. (2013) performed a similar analysis with SR at a global scale. The analysis relies on a Landsat-MODIS Consistency Checking System for evaluating Landsat SR products using near-simultaneous MODIS observations (Feng et al., 2012). Landsat-7 ETM + SR was directly compared to MODIS Terra SR while Landsat-5 TM SR was compared to 16-day Nadir Bidirectional Reflectance Distribution Function (BRDF) Adjusted Reflectance (NBAR) products (MCD43A4, Schaaf et al., 2002) to account for the significant sun-view geometries differences between any MODIS sensors and Landsat-5 TM. Feng et al. (2013) evaluated the MODIS-Landsat relative differences between 1.3% and 2.8% for Landsat-7 ETM + and between 2.2% and 3.5% for Landsat-5 TM. They concluded that systematic biases between Landsat and MODIS SWIR bands were likely due to bandwidth differences between the two systems. Finally, they identified calibration issues for some individual scenes.

These validation studies aimed at improving the quantification of the uncertainty of LEDAPS-derived SR Landsat data. However, Ju et al. (2012) and Maiersperger et al. (2013) validated the LEDAPS products over 107 and 26 US AERONET sites, respectively, whereas the network includes a total of about 500 sites globally distributed. Concerning the cross-comparison with MODIS, Maiersperger et al. (2013) used a very small number of Landsat-7 images. The analysis of Feng et al. (2013) is global and includes a large number of images. Nonetheless, Feng et al. (2013) used 16-day composite data NBAR products to compare to individual Landsat-5 TM scenes, assuming a low variation of the surface within the time-composite period. Moreover, the cross-comparison is carried out at 500 m, while MODIS 500 m resolution products are known to have a non-constant pixel footprint which increases with the view zenith angle (Wolfe, Roy, & Vermote, 1998 reported that the nominal 1 km pixel reaches 4.83 km along track on the scan edge). Similarly, Campagnolo and Montano (2014) have showed that the nominal 250 m resolution corresponds actually to an effective resolution varying from 292 m to 835 m. By applying a simple extrapolation of these findings, the effective resolution of the 500 m nominal resolution band may reach more than 1600 m. This reduces the accuracy of the relationship between the location of the grid cells and their input observations for large view angles. These spatial and temporal continuity assumptions may introduce some non-negligible noise in the cross-comparison.

Neither Maiersperger et al. (2013) nor Feng et al. (2013) included spectral adjustments when comparing TM/ETM + and MODIS SR. The MODIS has relatively narrow spectral bands compared to the Landsat TM and ETM + (Fig. 1). Although TM and ETM + relative spectral responses (RSR) are very similar, one can notice small differences of about 10 nm on blue, green, SWIR-1 and SWIR-2 bands. Feng et al. (2013) concluded that these represented a non-negligible source of discrepancy, particularly for the two Landsat SWIR bands. To build a relevant cross-comparison between Landsat and MODIS SR, it is thus needed to use a spectral adjustment method.

In this study, we propose to evaluate the consistency of LEDAPS-derived SR products from Landsat-5 TM and Landsat-7 ETM + images based on two approaches (AERONET validation and cross-comparison of SR Landsat-MODIS) described below.

First, we compare the SR truth computed across 489 AERONET globally-distributed sites by following the methodology introduced by Vermote and Kotchenova (2008). The main difference with Maiersperger et al. (2013) is the use of SR as the target variable to evaluate the uncertainty instead of the AOT. This difference is essential since SR is the final product delivered to the users. The current approach is identical to Ju et al. (2012), but we did not restrict the comparison to US sites.

Second, we designed an automated evaluation methodology to cross-compare Landsat-5 TM and Landsat-7 ETM + SR data with MODIS SR data using directional and spectral adjustments. The directional adjustment needed to correct for surface reflectance anisotropy (called hereafter BRDF adjustment) of the MODIS data is based on the VJB method (Vermote, Justice, & Breon, 2009). The method relies on the Ross-Thick-Li-Sparse BRDF model (Li, X.W. and Strahler, A.H., 1992, Lucht, W., 1998), similarly to the MCD43A4 MODIS NBAR products used by Feng et al. (2013), and includes the hot-spot as modeled by Maignan, Breon, and Lacaze (2004). Nonetheless, on the opposite of the MCD43A4 products, the VJB method does not rely on the 16-day SR stability assumption (Schaaf et al., 2002) since it operates with daily instantaneous MODIS SR products. The method supports two simplification hypotheses: (i) volumetric and geometric kernels are proportional to isotropic kernels using two coefficients, R and V, respectively, (ii) R and V are linearly related to the NDVI. In other words, this means that the BRDF shape varies per pixel according to vegetation cover variation, i.e. NDVI variation. BRDF have indeed been shown to be significantly different for bare soil and vegetated surfaces, because vegetated surfaces show higher anisotropy than bare soil do (Bacour & Breon, 2005). Based on inversion method introduced by Vermote et al. (2009), a database of the VJB coefficients have been retrieved globally using the very stable coarse spatial resolution (0.05°, called hereafter Climate Modeling Grid, CMG, resolution) SR MODIS products derived from spatially-averaged MODIS SR. The CMG resolution was selected in order to: (i) reduce the MODIS geometric error (discussed previously), and (ii) ensure a year-to-year land-cover-land-use stability through time (Becker-Reshef, Vermote, Lindeman, & Justice, 2010). The analysis in this paper, which is based on this available VJB coefficient database, is thus performed at the same CMG resolution, meaning that the 30 m Landsat pixels are aggregated over the CMG. The spectral adjustment of the MODIS data is processed using an innovative method based on an artificial neural network trained with a PROSAIL simulated SR database. For this analysis we used MODIS data issued from both Terra and Aqua platforms, since the BRDF adjustment method accounts for differences of sun-view geometry between MODIS and Landsat sensors.

Section snippets

AERONET data

The AERONET (Aerosol Robotic Network) project is a global network of about 500 ground-based instruments dedicated to the continuous measurement of aerosol properties (Holben et al., 1998). The project, initiated by NASA in the 1990s, relies on the standardization of instruments and procedures: measurement protocols, data processing (cloud screening, inversion techniques, etc.) and calibration procedures. The AERONET standard instrument routinely performs direct sun measurements with a 1.2°

AERONET validation

The first evaluation approach was based on 489 AERONET sites, where 7422 in situ measurements timely matched a Landsat scene (Fig. 2 bottom). We restricted the analysis to the MODIS-era period (2000–2013). Similarly to Ju et al. (2012), the analysis consisted in comparing reference SR (using AERONET measurements) to LEDAPS SR. We eliminated the pixels classified in LEDAPS quality assurance layers as: (i) cloud, (ii) cloud-shadow, (iii) water, (iv) snow and (v) with an atmosphere opacity higher

Validation of SR over AERONET sites

The surface reflectance data used for the comparison were selected on 10 km squares centered on the 489 AERONET sites from 2000 to 2013 (Fig. 2, top). Data were filtered from cloud, cloud-shadow, snow and water covers and SLC-off stripes of Landsat-7 ETM +. We retained a total of 259 and 298 million 30-m resolution pixels for Landsat-5 TM (2860 10 × 10 km subsets) and Landsat-7 ETM + (4562 10 × 10 km subsets), respectively. The results are shown in Fig. 4 for each of the 6 spectral bands of Landsat-5 TM

Conclusion

In this paper, we evaluated the performances of the LEDAPS surface reflectance products over land. Two evaluation approaches were used. The first approach relies on the globally distributed AERONET sites, using in situ atmosphere measurements to perform the atmospheric corrections and comparing the results to LEDAPS products. The same methodology has been described and discussed in the past by Ju et al. (2012) for Landsat sensors, but they used a limited number of AERONET sites (US only). The

Acknowledgments

The authors would like to thank the USGS ESPA On-Demand System (http://espa.cr.usgs.gov/) for processing the Landsat surface reflectance products. We acknowledge the anonymous reviewers for their very constructive comments. We also want to acknowledge Melanie Rosenberg for reviewing the English of this manuscript.

References (57)

  • S. Jacquemoud et al.

    PROSPECT plus SAIL models: A review of use for vegetation characterization

    Remote Sensing of Environment

    (2009)
  • J. Ju et al.

    Continental-scale validation of MODIS-based and LEDAPS Landsat ETM plus atmospheric correction methods

    Remote Sensing of Environment

    (2012)
  • V.M. Krasnopolsky et al.

    Some neural network applications in environmental sciences. Part II: Advancing computational efficiency in environmental numerical models

    Neural Networks

    (2003)
  • W.J.D. van Leeuwen et al.

    Multi-sensor NDVI data continuity: Uncertainties and implications for vegetation monitoring applications

    Remote Sensing of Environment

    (2006)
  • T.K. Maiersperger et al.

    Characterizing LEDAPS surface reflectance products by comparisons with AERONET, field spectrometer, and MODIS data

    Remote Sensing of Environment

    (2013)
  • F. Maignan et al.

    Bidirectional reflectance of earth targets: Evaluation of analytical models using a large set of spaceborne measurements with emphasis on the hot spot

    Remote Sensing of Environment

    (2004)
  • B.L. Markham et al.

    Forty-year calibrated record of earth-reflected radiance from Landsat: A review

    Remote Sensing of Environment

    (2012)
  • T. Miura et al.

    An empirical investigation of cross-sensor relationships of NDVI and red/near-infrared reflectance using EO-1 Hyperion data

    Remote Sensing of Environment

    (2006)
  • D.P. Roy et al.

    Conterminous United States demonstration and characterization of MODIS-based Landsat ETM plus atmospheric correction

    Remote Sensing of Environment

    (2014)
  • D.P. Roy et al.

    Landsat-8: Science and product vision for terrestrial global change research

    Remote Sensing of Environment

    (2014)
  • D.P. Roy et al.

    The MODIS land product quality assessment approach

    Remote Sensing of Environment

    (2002)
  • C.B. Schaaf et al.

    First operational BRDF, albedo nadir reflectance products from MODIS

    Remote Sensing of Environment

    (2002)
  • M.D. Steven et al.

    Intercalibration of vegetation indices from different sensor systems

    Remote Sensing of Environment

    (2003)
  • P.M. Teillet et al.

    Impacts of spectral band difference effects on radiometric cross-calibration between satellite sensors in the solar-reflective spectral domain

    Remote Sensing of Environment

    (2007)
  • A. Verger et al.

    Optimal modalities for radiative transfer-neural network estimation of canopy biophysical characteristics: Evaluation over an agricultural area with CHRIS/PROBA observations

    Remote Sensing of Environment

    (2011)
  • E.F. Vermote et al.

    Atmospheric correction of MODIS data in the visible to middle infrared: First results

    Remote Sensing of Environment

    (2002)
  • P.M. Atkinson et al.

    Introduction neural networks in remote sensing

    International Journal of Remote Sensing

    (1997)
  • F. Baret et al.

    Satellite derived leaf area index derived from SPOT time series in the ADAM project

  • Cited by (168)

    View all citing articles on Scopus
    View full text