Sie sind auf Seite 1von 4

# AN ILLUMINATION CORRECTION ALGORITHM ON LANDSAT-TM DATA

Bin Tan1,2,*, Robert Wolfe1, Jeffrey Masek1, Feng Gao1,2, Eric F. Vermote1,3
1 NASA Goddard Space Flight Center, Greenbelt, MD, USA. ERT, 10810 Guilford Road, Suite 105, Annapolis Junction, MD USA. 3 Department of Geography, University of Maryland, College Park, MD, USA. * Email: bin.tan@nasa.gov (corresponding author) 2

ABSTRACT
In this paper, a new illumination correction model, the rotation model, is introduced. The model is based on the empirical correlation between reflectance and the illumination condition (IL). The model eliminates the dependency of reflectance on IL through rotating the data in IL-reflectance space. This model is compared with widely used cosine model and C model over a sample forest region. We found that the newly developed rotation model consistently performs best on both atmospheric uncorrected and corrected Landsat images.
Index Terms Landsat, illumination correction, change detection, LEDAPS

to compare the performances of this model and traditional cosine correction and C correction models. We test them on both atmospheric un-corrected (topof-atmosphere reflectance, TOA) and corrected (topof-canopy reflectance, TOC) Landsat images.

2. METHODOLOGY
The solar incidence angle, or Illumination Condition (IL), is the basic of all correction models for compensating the reflectance:

## IL = cos Z cos S + sin Z sin S cos(z S )

Where Z is the solar zenith angle, S is the slope angle. z is the solar azimuth angle, and S is the aspect angle of the incline surface. 2.1 Traditional Models Two illumination corrections models are widely cos Z used, the cosine model L H = Li ( ) and the C IL , cos Z + c model L H ( ) = Li ( ) [5]. Where LH is IL + c the reflectance on flat surface, Li is the reflectance on incline surface, c is the ratio of the slope and intercept of one linear regression:

1. INTRODUCTION
Illumination correction, also known as topographic correction or topographic normalization, refers to the compensation of the solar irradiance to minimize the variability of observed reflectance for similar targets due to topography and BRDF effects. This is an important step in pre-processing high-resolution remote sensing data for forest change detection studies. Varying illumination conditions lead to significant changes in the spectral characteristics of a pixel, even in the absence of variations in land cover type or condition. The goal of this study is to produce illumination corrected Landsat-TM data with the quality needed for more accurate forest change detection [1][2] in the Landsat Ecosystem Disturbance Analysis Adaptive Processing System (LEDAPS) [3][4]. There are several existing illumination correction models for Landsat images. However, these models tend to overcorrect over low illuminated slopes. In this paper, we introduce an empirical based rotation-correction model. And a test area is selected

Li ( ) = a IL + b

(1)

Several studies have reported that the cosine model overcorrects the surface reflectance, especially in low IL regions [6][7][8]. The C model can avoid the overcorrection to some degree, but it is still significant in some low IL regions. Spectrally, both methods perform better in the near-infrared band than visible bands.

## 978-1-4244-9566-5/10/\$26.00 2010 IEEE

1964

2.2 Rotation Model The goal of illumination correction is to remove the dependency of the reflectance to IL. Following other models, we assume there is a linear relationship between reflectance and IL [8] [Eq. 1]. The following model is used to rotate the reflectance data in IL-reflectance space to remove the dependency of the reflectance on the IL:

LH ( ) = Li ( ) (a * IL + b)
where a and b is from Eq. 1.

(2)

We examine the performance of each model by two ways: (1) whether the corrected image preserves the spectral structure of the original image. Is there any significant overcorrection or undercorrection, which could lead difficulties to the subsequent studies? (2) Whether the corrected image increases the reflectance homogeneity of forest region, which means reducing the reflectance changes due to different illumination conditions. The test area is located in a forested region of Northern California (a subset of Landsat-7 scene over path 46 and row 32 archived on Oct 1st, 2002). The atmospheric corrected image is retrieved with 6S model [9]. All three illumination correction models are applied to both images.

## 3. RESULTS AND DISCUSSIONS

The original and corrected images are shown at Fig. 1. The RGB composite is using near-infrared, red, and green band respectively. In general, all three models can compensate the illumination variance due to terrain effects. However, the cosine model performs poorly on TOA data. The visual bands (blue, green, and red) are overcorrected (blue band is not shown here). The band with shorter wavelength has more overcorrection. This overcorrection is due to the ignorance on the effect of diffuse radiation. In the cosine model, the terrain effect is compensated cos Z . This could compensate with the coefficient IL TOC reflectance correctly, where the diffusion of the atmosphere has been removed, or at least minimized. However, this is not the case for TOA reflectance, where diffusion radiation is significant in visual bands. Theoretically, only the direct incident beam

Fig. 1. Landsat images of original data, corrected by cosine model, corrected by C model, and corrected by the rotation model, from top to bottom respectively. The left column is TOC results and the right column is TOA results. The RGB composite is using near-infrared, red, and green band respectively.

1965

cos Z should be corrected with the coefficient IL . The diffusion incident is not impacted by the incidence angle, which then should not be compensated. When cos Z coefficient IL is applied to TOA data, both direct and diffuse sources are actually compensated. This leads to the overcorrection.
The C model overcorrected some pixels with low IL in red band of TOC reflectance as shown in Fig 2 (the third image in the left column). The overcorrection comes from the leverage coefficient

when C is a negative value and the absolute value is close to the value of IL. This usually happens in low IL region and in the band whose reflectance has a weak correlation to IL, e.g. red band. Although the overcorrection only happens in TOC reflectance in this study, it should be noted that this could happen in TOA reflectance too. The rotation model does not suffer the overcorrection because there is no such leverage coefficient in the model. In our studies, the rotation model performs better than either the C model or the cosine correction on both TOA and TOC reflectance. Table 1 summarizes the standard deviation of each TOA and TOC Landsat band before and after illumination correction in the area showing in Fig 2. As expected, the standard deviation of corrected images is lower than the original images, meaning greater homogeneity. The corrected TOC reflectance from C model has a high variance in red band and band-5. The cosine model performs well on TOC reflectance. On TOA reflectance, the C model produces a more homogeneous result than the cosine model, especially on the blue band. For both TOA and TOC reflectance, the rotation model produced the best result with lowest standard deviation. 4. CONCLUSIONS This paper presents an illumination correction model for Landsat data. This model is based on the empirical relationship between reflectance and IL. We compared it with traditional cosine and C model. The test result suggests the newly developed rotation

## cos Z + c . The overcorrection happens IL + c

Fig. 2. Zoom-in images of Fig.1 over low IL region. The subset is marked with the green box in the up-left image in figure 1.

model performance consistently good on both TOA and TOC Landsat data. So we can conclude that the rotation model is more applicable to Landsat data in forest areas. Further research should be done in the following aspects:

1966

Validate the correction results via different methods, including comparing with field measurements, and other satellite data which were taken under different sun angles etc. Apply the rotation model on other land cover types. Since the terrain effect is vegetation dependent [8], a separate correlation should be established for each land cover type. Apply change detection algorithms on corrected data to assess whether the illumination corrected data could improve the change detection accuracy.

Table 1. Standard deviation of TOC (upper part) and TOA (lower part) reflectance and illumination corrected reflectance by each model.
Band 1 2 3 4 5 7 Band 1 2 3 4 5 7 Original 0.00492 0.01214 0.01844 0.05100 0.05254 0.02280 Original 0.00365 0.00989 0.01593 0.04680 0.04921 0.02280 Cosine Model 0.00471 0.01112 0.01769 0.04327 0.04873 0.02162 Cosine Model 0.01268 0.00991 0.01492 0.03975 0.04621 0.02182 C Model 0.00510 0.01110 0.01883 0.04389 0.05248 0.01912 C Model 0.00321 0.00885 0.01496 0.03987 0.04592 0.01962 Rotation Model 0.00426 0.01062 0.01701 0.03800 0.04582 0.01888 Rotation Model 0.00317 0.00877 0.01488 0.03526 0.04364 0.01933

5. REFERENCES
[1] Masek, J. G., S. Goward, W. Cohen, R. Wolfe, C.
Huang and F. Hall, Assessing North American Forest Disturbance from the Landsat Archive, IGARSS 2007: International Geosci. and Remote Sens. Symposium, pp. 5294-5297, July 2007. [2] Masek, J. G., C. Q. Huang, R. Wolfe, W. Cohen, F. Hall, J. Kutler, P. Nelson, North American forest disturbance mapped from a decadal Landsat record, Remote Sens. Environ., 112(6):2914-2926, 2008. [3] R. Wolfe, J. Masek, N. Saleous, F. Hall, LEDAPS: mapping North American disturbance from the Landsat record, IGARSS 2004: International Geosci. and Remote Sens. Symposium, pp. 1-4, July 2004. [4] Masek, J. G., E. F. Vermote, N. E. Saleous, R. Wolfe, F. G. Hall, K. F. Huemmrich, F. Gao, J. Kutler, and T.-K. Lim, A Landsat Surface Reflectance Dataset for North America, 1990-2000, IEEE Trans. Geosci. Remote Sens. Letters, 3 (1):68-72, 2006. [5] P. M. Teillet, B. Guindon, And D. G. Goodenough, On the slope-aspect correction of multispectral scanner data, Can. J. Remote Sens., vol. 8, pp. 84106, 1982. [6] C. R. Duguay and E. F. LeDrew, Estimating surface reflectance and albedo from Landsat-5 TM over rugged terrain, Photogramm. Eng. Remote Sens., vol. 58, pp. 551-558, 1992. [7] P. Meyer, K. I. Itten, T. Kellenbenberger, S. Sandmeier, and R. Sandmeier, Radiometric corrections of topographically induced effects on Landsat TM data in an alpine environment, ISPRS J. Photogramm. Remote Sens., vol. 48, pp. 17-28, 1993.

[8] D. Riano, E. Chuvieco, J. Salas, and I. Aguado, Assessment of difference topographic corrections in Landsat-TM data for mapping vegetation types, IEEE Trans. Geosci. Remote Sensing, vol 41, pp. 1056-1061, 2003. [9] E. F. Vermote, N. El Saleous, C. O. Justice, Y. J. Kaufman, J. L. Privette, et al., Atmospheric correction of visible to middle-infrared EOS-MODIS data over land surfaces: Background, operational algorithm, and validation, J. Geophy. Research, 102, 17131-17141, 1997.

1967