The empirical line method for the atmospheric correction of IKONOS imagery

INT. J. REMOTE SENSING, 2003, VOL. 24, NO. 5, 1143 1150 The empirical line method for the atmospheric correction of IKONOS imagery E. KARPOUZLI* and T. MALTHUS Department of Geography, University of Edinburgh, Drummond Street, Edinburgh EH8 9XP, UK (Received 9 July 2001; in final form 13 August 2002) Abstract. The empirical line method is an atmospheric correction technique that provides an alternative to radiative transfer modelling approaches. It offers a relatively simple means of surface reflectance calibration, providing that a series of invariant-in-time calibration target measurements are available. This technique has been applied with variable success to both airborne data and coarser spatial resolution satellite sensor data. However, with the advent of higher spatial resolution space-borne sensors there is a need to re-evaluate its potential. The empirical line method was tested for correcting multispectral IKONOS imagery acquired over the tropical island of San Andres, Colombia. The high spatial resolution (4 m) of the data made it possible to identify a number of homogeneous targets with a range of reflectances that were used for the calibration. Coefficients of determination of the prediction equations observed were large, ranging from 0.96 0.99 for each of the four wavebands. An accuracy assessment was performed using a set of independent targets. It demonstrated that the empirical line method can be applied to correct such imagery with accurate results. 1. Introduction The effects of atmospheric scattering and absorption must be removed from satellite remotely sensed data if such images are to be used quantitatively, either on their own or within a temporal data set to detect environmental change. Atmospheric effects add to or diminish true ground leaving radiance, and act variably across the optical spectrum. Their removal is particularly important for marine applications due to the proportionately greater contribution to at-sensor received radiance over water targets (e.g. Gordon 1997, Gordon et al. 1997). A number of methods exist to account for the influence of the atmosphere on sensor-recorded radiance, including the dark pixel or histogram method (e.g. Chavez 1996), covariance matrix method and radiative transfer models like MODTRAN (e.g. Ferrier 1995), EXACT (Popp 1995), and 6S (Vermote et al. 1997). Whilst radiative transfer models may be the preferred method, their input parameters often prove difficult to obtain accurately, especially when dealing with historical datasets. Other methods include empirical relationships between radiance and reflectance, * e-mail: e.karpouzli@hw.ac.uk e-mail: tjm@geo.ed.ac.uk International Journal of Remote Sensing ISSN 0143-1161 print/issn 1366-5901 online 2003 Taylor & Francis Ltd http://www.tandf.co.uk/journals DOI: 10.1080/0143116021000026779

1144 E. Karpouzli and T. Malthus such as the empirical line method (e.g. Ferrier 1995, Smith and Milton 1999). However, to date this method has been applied only to the relatively short atmospheric thicknesses of airborne data for the principal reason that it requires the identification of at least two homogeneous targets of contrasting reflectances that are large enough to be resolved (Perry et al. 2000). It is difficult to find such areas at spatial resolutions suitable for calibrating spaceborne images with pixel sizes of tens of metres such as Landsat Thematic Mapper (TM) and Système Probatoire d Observation de la Terre (SPOT). However, with the advent of satellite sensors capable of measurement at a much higher spatial resolution (e.g. 1 to 5 m pixel size, such as IKONOS and QuickBird2) there is now the potential for applying this technique to spaceborne data. A test of this method for the longer atmospheric pathlengths of satellite sensor data is warranted. In this research, the empirical line method was tested for its ability to atmospherically correct an IKONOS image acquired over the Caribbean island of San Andres. This method offers a relatively simple means of surface reflectance calibration, provided that a series of invariant-in-time calibration target measurements are available (Teillet et al. 1990, Smith and Milton 1999). Although it can be performed using just two calibration targets, the use of more targets allows the parameters of the relationship between at-sensor radiance and at-surface reflectance to be estimated with greater confidence; there is evidence that acceptable results can be obtained using four calibration targets or more (Price et al. 1995, Smith and Milton 1999). 2. Methods An IKONOS multispectral (4 m) image was acquired over the island of San Andres (Colombia) in the Western Caribbean Sea, on 9 September 2000, at 10:33 am local time (figure 1). These data were already radiometrically corrected from digital numbers (DN) to in-band radiance physical units (mw cm 2 sr 1). Subsequently the image was geometrically corrected using Ground Control Points (average RMS: 0.87 m). The high spatial resolution of the data made it possible to identify several suitable targets on the island to implement the empirical line method. In situ reflectance measurements were made between 1 and 10 September 2000. Measurements were performed between 10:00 and 13:00 hours local time under conditions of scattered cumulus cloud or clear skies. Nine calibration targets were used (figure 1), and for accuracy assessment purposes a further five validation targets were identified and measured (table 2). The criteria for the selection of both calibration and validation targets, following Smith and Milton (1999), and Che and Price (1992), were: $ Areas as spectrally homogeneous as possible, preferably Lambertian and horizontal. $ Areas of a range of reflectances and preferably devoid of vegetation. $ Areas of a size at least three times the IKONOS pixel size, i.e. at least 12 12 m. The last criterion was not always easy to meet. In cases of elongated targets like roads or beaches one dimension was typically less than three times the IKONOS pixel size (i.e. 8 10 m wide). The targets selected ranged from dark reflectance targets such as black asphalt and deep water, to bright reflective ones like sand, light soil, and concrete (table 1). To increase the sample a number of intermediate but less ideal calibration targets were included, such as areas of different grass species or grass under different

Remote Sensing L etters 1145 Figure 1. IKONOS multispectral image of the island of San Andres, Colombia, and position of calibration targets used for its atmospheric correction. management practices. Since the ground measurements were made within 9 days of the acquisition of the IKONOS image, the vegetated targets were believed to be invariant in time (CORALINA Institute, San Andres, personal communication, 2000). Reflectance measurements for each of the calibration/validation targets were made using a GER 1500@ spectroradiometer, (300 1100 nm, nominal dispersion 1.5 nm, spectral resolution: 3 nm). A single sensor head was used fitted with a 15 lens giving a footprint of approximately 16 cm diameter. References to incident

1146 Table 1. E. Karpouzli and T. Malthus Description of targets used for the atmospheric correction of the IKONOS image, and mean coefficients of variation (COV) between spectra averaged. Calibration target COV identification Description (%) 1 White coraline sand on beach 8.86 2 White coraline sand on beach 6.62 3 Light coloured soil ~50 60% sand at football ground 7.50 4 Darker soil ~40% sand at football ground 6.35 5 Red coloured soil at baseball ground 11.52 6 New asphalt on main road outside airport, no road markings 12.16 7 80% cover irrigated short turf at baseball pitch, regularly cut, 11.39 dominated by clover 8 80% cover tall grass, non-irrigated part of baseball pitch, uncut 21.00 9 Over deep water (>90 m depth) 15.86 Table 2. Description of targets used for validation of the atmospheric correction of the IKONOS image and mean coefficients of variation (COV) between spectra averaged. Validation target COV identification Description (%) 1 95% cover short dense turf in handball pitch 21.05 2 Old asphalt on main road, no road markings 15.56 3 70% cover short grass and occasional sedges in handball pitch 12.40 4 White coraline sand on the beach 8.86 5 New road asphalt, some bright road markings 9.38 irradiance over a calibrated SpectralonTM panel were obtained for every 3 10 target measurements, the frequency depending on the degree and change of cloud cover at the time of the measurement. Between 15 and 60 spectra were taken for each target depending on the variation visible within the target. Although an effort was made to identify visibly homogeneous targets, some heterogeneity existed, especially within vegetation targets (e.g. in density). In such cases, a larger number of spectra were measured and averaged to account for their spatial variation (Milton et al. 1997). A Garmin 100 SRVY II GPS, operated in non-differential mode (Selective Availability turned off ), was used in the averaging setting for at least 10 min to estimate the centre position of each target (Garmin 1993). The spectral data were processed to absolute reflectance. From the averaged reflectance spectra for each target the reflectance values for the IKONOS bandwidths were calculated using filter functions based on the sensor response curves provided by Space Imaging Inc., Thorton, Colorado. 3. Results and discussion The averaged reflectance spectra for each of the targets show a good range of measured reflectances across all wavelengths (figure 2). Coefficients of variation for the calibration targets ranged from approximately 6% to 21%, with the tall grass target showing the greatest heterogeneity (table 1)

Remote Sensing L etters 1147 Table 3. Comparison of validation target IKONOS band reflectances (%), predicted from the empirical line relationships in figure 3, and their calculated actual band reflectances. IKONOS band Validation Blue Green target identification Predicted Actual DiVerence Predicted Actual DiVerence 1 6.3 6.3 0 10.2 9.5 0.6 2 8.3 5.9 2.4 8.2 6.2 2 3 8.1 6.4 1.8 12 10 2 4 49 48.1 0.9 56.5 55.8 0.7 5 7.1 8.5 1.4 7.1 9 1.8 IKONOS band Validation Red NIR target identification Predicted Actual DiVerence Predicted Actual DiVerence 1 9.8 8.9 0.9 39.4 39 0.4 2 9 6.4 2.6 8.5 6.6 1.9 3 13.6 10.8 2.7 35.9 36.1 0.3 4 66.6 65.7 0.9 68.6 71.1 2.4 5 8.4 9.2 0.8 8.1 9.6 1.5 Figure 2. Averaged reflectance spectra for the San Andres calibration targets. Target numbers correspond to target descriptions in table 1. Radiance values recorded by the sensor were extracted from the image by averaging the values of 1 3 pure pixels associated with each calibration target (Smith and Milton 1999). These values in each band were plotted against the corresponding ground reflectances for the calibration targets, and linear regression

1148 E. Karpouzli and T. Malthus was used to derive a set of four prediction equations for the image, one for each waveband (figure 3). The coefficients of determination for the three visible bands were all greater than 0.97, and for the near-infrared band was 0.96. In general, the relationships found were near linear, which agrees with the findings of Stow et al. (1996) who used four calibration targets covering a wide range of reflectance values, to correct multispectral imagery from an airborne digital camera. The interception of the calibration lines with the x-axis represents the atmospheric path radiance (Smith and Milton 1999). As can be observed from the individual plots the path radiance was greatest for the blue band and decreased with increasing band wavelength, due to greater scattering by the atmosphere. The slopes of the calibration lines represent the atmospheric attenuation. The prediction equations were applied to each band of the original geometrically corrected image of the island, to produce a new dataset calibrated to absolute reflectance units. To assess the error in the empirical relationships derived, the five validation targets were used. The reflectance values for each target for each IKONOS band were calculated from the ground spectral measurements in the same way as for the calibration targets, and image reflectances were derived for corresponding pixels from the calibrated image (1 3 pure pixels). Overall, estimated reflectances were within ±3 reflectance units of actual reflectances. 4. Conclusions Although the validation and calibration targets used in the study were similar in nature, the large correlation coefficients observed between at-sensor radiance and ground reflectance for the four IKONOS wavebands and the independent error assessment demonstrates that the empirical line method can be applied to correct IKONOS imagery with highly satisfactory results. The increased spatial resolution of this sensor made it possible to identify a large number of natural targets that were sufficiently homogeneous and larger then 2 to 4 times that of the pixel size. When such empirical methods are applied it is essential to ensure that certain conditions are met, especially if there are no validation data to assess how well the empirical model has performed. The calibration targets used must be selected carefully to cover the whole range of reflectances, to be of an appropriate size in respect to the pixel size of the imagery and have as much as possible near-lambertian properties. To ensure that a wide range of reflectances are measured and to reduce extrapolation beyond the calibration data, preliminary spectral measurements should be made before final selection of the targets. If ground measurements cannot be made concurrent with the image acquisition then the targets should be spectrally stable in time. The empirical line method assumes that the effects of the atmosphere are uniform across the image, which is often not the case and that there are no differences in illumination across the image, i.e. due to shadows from topography or clouds (Smith and Milton 1999). For this study area the calibration targets were all within the scene to be corrected and the small coverage of the imagery suggests that spatial variation of the atmosphere would be small. Whilst some cloud shadowing was present, topographic shadowing was less of an issue in this low-lying terrain. Unlike Ferrier (1995) who reported disappointing results when using the empirical line method to atmospherically correct AVIRIS imagery (attributed to inaccuracies in locating the targets on the image), both Price et al. (1995) and Smith and Milton

Remote Sensing L etters 1149 Figure 3. Regression lines and prediction equations using nine calibration targets for each of the IKONOS bands.

1150 Remote Sensing L etters (1999) reported accurate results when correcting CASI imagery, with errors of only a few per cent with two to four calibration targets. The results reported here suggest that this error can be reduced further when using a greater number of calibration targets. Ferrier (1995) suggested that target height differences (of about 110 m) were possible reasons for errors being introduced. In the case of our dataset all targets were at sea level, which may have led to the increased correlations derived. Acknowledgments This research was supported by the Darwin Initiative, Onassis Foundation, Carnegie Trust, the University of Edinburgh, Reef UK and the Moray Development Fund. It was conducted in collaboration with CORALINA in San Andres. The assistance of several colleagues during fieldwork is gratefully acknowledged. The spectroradiometer was obtained on loan from the Natural Environment Research Council Equipment Pool for Field Spectroscopy, UK. References CHAVEZ, P. S., 1996, Image-based atmospheric corrections revisited and improved. Photogrammetric Engineering and Remote Sensing, 62, 1025 1036. CHE, N., and PRICE, J. C., 1992, Survey of radiometric calibration results and methods for visible and near infrared channels of NOAA 7, 9 and 11 AVHRR. Remote Sensing of Environment, 41, 19 27. FERRIER, G., 1995, Evaluation of apparent surface reflectance estimation methodologies. International Journal of Remote Sensing, 16, 2291 2297. GARMIN, 1993, GPS 100 SRVY II Owner s manual. Garmin, Lenexa, Kansas, USA. GORDON, H. R., 1997, Atmospheric correction of ocean colour imagery in the Earth Observing System era. Journal of Geophysical Research-Atmospheres, 102, 17 081 17 106. GORDON, H. R., ZHANG, T. M., HE, F., and DING, K. Y., 1997, Effects of stratospheric aerosols and thin cirrus clouds on the atmospheric correction of ocean colour imagery: simulations. Applied Optics, 36, 682 697. MILTON, E. J., LAWLESS, K. P., ROBERTS, A., and FRANKLIN, S. E., 1997, The effects of unresolved scene elements on the spectral response of calibration targets: an example. Canadian Journal of Remote Sensing, 23, 252 256. PERRY, E.M.,WARNER, T., and FOOTES, P., 2000, Comparison of atmospheric modelling versus empirical line fitting for mosaicking HYDICE imagery. International Journal of Remote Sensing, 21, 799 803. POPP, T., 1995, Correcting atmospheric masking to retrieve the spectral albedo of land surfaces from satellite measurements. International Journal of Remote Sensing, 16, 3483 3508. PRICE, J. C., ANGER, C. D., and MAH, S., 1995, Preliminary evaluation of CASI preprocessing techniques. 17th Canadian Symposium on Remote Sensing, Saskatoon, Canada (Canadian Remote Sensing Society), 12 15 June 1995, 694 697. SMITH, G. M., and MILTON, E. J., 1999, The use of the empirical line method to calibrate remotely sensed data to reflectance. International Journal of Remote Sensing, 20, 2653 2662. STOW, D., HOPE, A., HGUGEN, A. T., PHINN, S., and BENKELMAN, C. A., 1996, Monitoring detailed land surface changes using an airborne multispectral digital camera system. IEEE T ransactions on Geoscience and Remote Sensing, 34, 1191 1203. TEILLET, P. M., SLATER, P. N., DING, Y., SANTOR, R. P., JACKSON, R. D., and MORAN, M. S., 1990, Three methods for the absolute calibration of the NOAA AVHRR sensors in-flight. Remote Sensing of Environment, 31, 105 120. VERMOTE, E. F., TANRE, D., Deuze, J. L., Herman, M., and Morcrett, J. J., 1997, Second simulation of the satellite signal in the solar spectrum, 6S: an overview. IEEE T ranscactions on Geoscience and Remote Sensing, 35, 675 686.