Assessment of Atmospheric Correction Methods for Landsat TM Data Applicable to Amazon Basin Research Dengsheng Lu, Paul Mausel (Department of Geography, Geology and Anthropology, Indiana State University)

Eduardo Brondizio and Emilo Moran (ACT/Department of Anthropology, Indiana University)

0

5

10

15

20

25

30

1 2 3 4 5 6

surf

ace

refle

ctan

ce o

f for

est (

%)

DATE1 DATE2GAIN1 GAIN2COST1 COST2

0

5

10

15

20

25

30

35

40

1 2 3 4 5 6Surfac

e refle

ctance

of ba

re soil

(%)

DATE1 DATE2GAIN1 GAIN2COST1 COST2

0

10

20

30

40

50

60

70

80

1 2 3 4 5 6

TM D

N va

lues o

f fores

t

1991 TM image 1988 TM image

1987 TM image 1985 TM image

0

5

10

15

20

25

30

35

1 2 3 4 5 6

Surf

ace

refle

ctan

ce o

f for

est (

%)

1991 TM image1988 TM image1987 TM image1985 TM image

0

10

20

30

40

50

60

70

80

90

100

1 2 3 4 5 6 7 8 9 10 11 12

TM bands {raw band 1-6 (TM1-5,7) in DN value, corrected band 7-12 (TM1-5,7) in % }

Spect

ral Re

sponse

A013-93 SS2

A012-93 SS2

A001-93 SS3

A002-93 SS3

A003-93 LianaForestA008-93 UplandForestA005-93 SS3

A010-93 SS1

A001-92 SS3

A011-93 SS2

A006-93 SS2

A007-93 SS2

A009-93 SS1

A014-93 SS2

Abstract

Atmospheric correction is an important preprocessing step in remote sensing digital image analysis, especially when multi-temporal data are applied to extract such information as biomass and land use/land cover features and their changes. Models, such as 6S, LOWTRAN, and image-based DOS (dark object subtraction), have been developed for atmospheric correction. Which model is appropriate depends on the requirements of the research application. The 6S model has advantages over other models in accuracy and applicability in a variety of circumstances, but it is complex and requires detailed atmospheric data acquired at the time of satellite overflight. DOS models overcome the implementation problems of the 6S model since it is based on the image measurement and is easy to apply. In this paper, two methods using theoretical spectral radiance and image acquisition date respectively were used to convert Thematic Mapper (TM) DN values to at-satellite radiance and three image-based models (apparent reflectance, DOS, and improved DOS) were employed to convert at-satellite radiance to surface reflectance in an Altamira, Brazil study area of the Amazon Basin. Interpretation of spectral graphs derived from different image-based models were conducted. The improved image-based DOS model was found to be the best technique for correcting atmospheric effects in this research. The correction of path radiance and atmospheric transmittance can greatly improves the image quality so that the surface reflectance of the different images with spatial and temporal scales have the consistency and comparability needed for good analysis.

Introduction

The Indiana University/Indiana State university (IU/ISU/INPE) LBA-Ecology project “Human Dimensions of Amazonia: Forest Regeneration and Landscape Structure” requires remotely sensed digital data to permit multitemporal, multisensor, across scene, enhanced data development, and selected biophysical modeling to be comparable and accurate. Currently, more than 30 Landsat TM scenes in seven Amazon study areas are being used in the IU/ISU research to develop land use/land cover (LULC) inventories, LULC change with a focus on succession, and selected biophysical modeling such as biomass. These research foci benefit from or absolutely require that the remotely sensed data used has been atmospherically corrected or calibrated. This paper explores the calibration methods which are most appropriate for the IU/ISU/INPE LBA research.

Study Area and Data Characteristics

The study area used to test calibration is located near the Altamira, Brazil (TM path and rows: 226 and 62). It is identified in red by number 1 in figure 1. Other numbers show the location of other study areas to which atmospheric calibration will be applied initially. This region is dominated by dense and liana forests, large areas with various stages of secondary vegetation, pasture lands, and annual or perennial crops. The study area presents an ideal laboratory to study land cover trajectories and biomass dynamics representative of the eastern Amazon. Four different dates of Landsat TM imagery were used in this calibration study. The image acquisition date, sun elevation angle, etc. are listed in table 1. This multitemporal data set used is representative of a variety of data quality conditions found in this region.

Methods

Three image-based models (apparent reflectance, DOS, and improved DOS) were used to atmospheric correction. Table 2 provides the information about the methods used in this research.

The first step of the image-based atmospheric correction method is to convert remotely sensed DN values to at-satellite radiance based on the gain and bias for each band, which was provided from the image header file.

L.sensor = Gain * DN + Bias ---------------------------------------- (1)

However, for most historical Landsat imagery of the Altamira area in this research, the gain and bias for each band are not available. The only available information that can be obtained is the image acquisition date. Two methods for converting TM DN values to at-satellite radiance were employed in this research. The first method is based on maximal and minimal spectral radiance values for each band.

L.sensor = ( L max – L min )/L range * DN + L min -------------------(2)

The second method is based on the image acquisition date.

L sensor = ( DN – DSL )/G --------------------------------------- (3)

G =MF * Days + AF ---------------------------------------------- (4)

The first step eliminates or reduces the effects from the satellite sensor system, and the second step converts the apparent at-satellite radiance to surface reflectance which involves the correction of effects caused by both solar angle and the atmosphere. The earth-sun distance can be obtained from the Astronomical Almanac according to the image acquisition date. The sun zenith angle can be obtained from from the image header file or from mathematical calculation based on the image acquisition date and time, and the longitude and latitude of the study area Three techniques based on the image information were used in this research. The first method (apparent reflectance model) corrects the effects caused by the solar radiance and sun zenith angle while ignoring the effects caused by atmospheric scattering and absorption. This method is very simple and easy to apply because it does not require in-situ field measurements.

R = PI * D2 * L sensor/(Esun * COS())---------------------------- (5)

The second method (DOS model) is also strictly an image-based procedure. It corrects for the effects caused by sun zenith angle, solar radiance, and atmospheric scattering, but can not correct the atmospheric absorption.

R = PI * D2 * (L sensor – L.haze)/(Esun * COS()) ----------------- (6)

The third method (improved DOS model) has the all functions of the above two models and also takes into account the atmospheric multiplicative transmittance components.

R = PI * (L sensor – L.haze)/(TAUv * Esun * COS() * TAUz )-- (7)

The TAUv and TAUz can be estimated from optical thickness. Chavez found two methods to estimate the atmospheric transmittance. One method uses the cosine of the solar zenith angle for TAUz which is called the COST model. Another method uses default TAUz values which are the average for each spectral band derived from the radiative transfer code. The TAUv is equal to 1 because the viewing angle for Landsat TM images is zero. Table 3 lists the constants that were used in the models.

Which method is most suitable for atmospheric correction in the Amazon Basin area? Comparison of spectral graphs of the same land cover, derived from different calibration models, can be used to analyze which calibration method is the most reasonable based on the spectral distribution in different wavelengths. The spectral graphs of invariant land cover from multi-temporal imageries, derived from the same calibration method, can be used to assess the reliability of this method because reflectance of the same land cover should have the same value on different image acquisition dates if atmospheric and radiometric effects are removed and environmental conditions are similar.

Results and Discussion

Figure 2 is a comparison of forest reflectance using different calibration methods. The reflectance of each band is the average value from the dense mature forests at different sites. It is assumed that large dense mature forests have a relatively stable surface reflectance at the spatial and temporal scales. The apparent reflectance models (DATE1 and GAIN1) resulted in high surface reflectance in the visible bands. This is because the apparent reflectance model only corrects for the effects caused by the sun angle, sun-earth distance, and the solar radiance. It ignores the effects of atmospheric scattering. The visible bands are mainly influenced by the atmospheric scattering which provide image additive effects. The DOS models (DATE2 and GAIN2) correct for the sun angle, sun-earth distance, solar radiance, and atmospheric scattering. The forest reflectance in the visible bands is very low due to the strong chlorophyll absorption. The improved DOS models (COST1 and COST2) has higher reflectance values than those obtained from the DOS model and the apparent reflectance model. In addition, the reflectance from COST1 has higher values than that from the COST2 due to the different method used in converting Landsat DN values to at-satellite radiance. In the near and middle infrared bands the improved DOS models have higher reflectance than from the other models used in this research. This is because the COST (COST1 and COST2) method reduces the effects caused by the atmospheric transmittance which decrease the surface reflectance. The DOS models and apparent reflectance models produced similar results in the near and middle infrared bands because atmospheric scattering is very weak within these wavelength. If the atmospheric scattering is ignored, the two models became identical. The image acquisition date approach resulted in higher surface reflectance than that using the theoretical maximal and minimal spectral radiance. Figure 3 is the surface reflectance of bare soils.This figure further confirms that using the image acquisition date in converting Landsat TM DN values

Figure 1: The Location of Selected Study Area

to at-satellite radiance can produce higher reflectance than that using the maximal and minimal radiance. In addition, the methods of using maximal and minimal radiance underestimated the surface reflectance, especially in the near and middle infrared wavelength. The improved DOS model can produce higher reflectance due to its capability of correcting for the atmospheric scattering and transmittance. The COST1 method provides the most reasonable results. Figure 4 compares of mature forest reflectance on different years using the COST1 method. This method provides stable and good results in these four dates of TM images. The 1991 image has a relative high surface reflectance compared with the other three images, especially in the visible wavelengths. Heavy haze exists in the 1991image and the visible bands also have stripping problems. The graphs indicate that the path radiance can not be ignored in the near infrared and middle infrared wavelength when heavy haze exists. Scenes with clouds indicate that there is more humidity in the atmosphere than average. High humidity lowers surface reflectance, especially in the near and middle infrared wavelength. Figure 4 shows that after calibration using COST1, all four dates have nearly identical spectral characteristics for forest. This result provides high quality multitemporal analysis to be conducted. Figure 5 shows the dense mature forest DN values of the raw TM data sets on different dates. The 1991 image has higher DN values in the visible bands than that of other three dates of images, but has somewhat lower DN values in the near and middle infrared bands compared with the clear weather 1985 image. This confirms that the heavy haze increases the DN values in visible bands and reduces the DN values in the near and middle infrared bands. Haze is an additive function that makes the DN values higher than that of clear weather in the shorter wavelengths. The atmosphere associated with clouds also influences the DN values, especially in the near and middle infrared wavelengths that decrease in the DN values due to atmospheric absorptions. Note how the uncalibrated spectral signature varies from date to date making many types of multitemporal analysis unreliable. Figure 6 compares 1991 TM raw and calibrated images after the stripping effects were removed for forest and secondary succession features. The quality of calibrated imagery has greatly improved, especially in visual bands, after removing the stripping effects.

TM L max L min MFl AFl DSLl Esun l TAUzTM 1 15.21 -0.15 -3.58E-05 1.376 2.523 195.29 0.70

TM 2 29.68 -0.28 -2.10E-05 0.737 2.417 182.74 0.78TM 3 20.43 -0.12 -1.04E-05 0.932 1.452 155.00 0.85TM 4 20.62 -0.15 -3.20E-06 1.075 1.854 104.08 0.91TM 5 2.72 -0.04 -2.64E-05 7.329 3.423 22.08 **TM 7 1.44 -0.02 -3.81E-04 16.020 2.633 7.50 **

Date method

Maximal and minimal radiance

Apparent reflectance model DATE1 GAIN1

Dark object model (DOS) DATE2 GAIN2

Improved image based DOS model

COST1 COST2

Methods used to convert at-satellite radiance to surface reflectance

Methods used to convert DN value to at-satellite radiance

 

Image acquisition date

Sun elevation angle (degree)

Cloudy conditions (%)

Aug. 4, 1985 48 Cloud free: 0%Jun. 23, 1987 45 Cloudy: 10%Jul. 11, 1988 46 Cloudy: 10%Jul. 20, 1991 45 Cloud free: 0%Note: Heavy haze exists in 1991 TM imagery and there is stripping in the visual bands.

Conclusions

There are two major conclusions in this research:

1) The improved DOS model, specifically COST1, is satisfactory to atmospherically calibrate all the multitemporal TM data currently used in the IU/ISU/INPE LBA Amazon research; 2) The quality of the DOS-based correction used is comparable to a good physically-based model, but it requires much less effort and cost to implement.

Note:SS1: Early Secondary Succession SS2: Intermediate Secondary SuccessionSS3: Advanced Secondary Succession

Table 3: Constants Used in the Models

Table 2: Methods Used in Research

Table 1: Image Data Used in Research

7 777

Figure 2: Comparison of Calibration Methods (1985 TM Image)

Figure 3: Comparison of Calibration Methods (1985 TM Image)

Figure 4: Comparison of Multidate Images (COST1 Method) Figure 5: Comparison of Raw TM Data

Figure 6: Comparison between Raw Data and Corrected Data on Selected Study Sites (1991 TM Image)