A method to detect and correct single-band missing pixels in Landsat TM and ETM+ data

Computers & Geosciences 34 (2008):445-455 Tobias Kuemmerle, Alexander Damm, and Patrick Hostert © 2007 Elsevier Ltd. All rights reserved doi:10.1016/j.cageo.2007.05.016 Received 2nd February 2006; revised 9th September 2006; accepted 8st May 2007

Abstract

Human land use is the main driver of terrestrial ecosystems change, and remote sensing is an important tool to monitor these changes. Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper (ETM+) images have been the most important data source to map land cover change, but image artifacts often hinder or even prohibit digital change detection. This paper addresses a group of image distortions that display erroneous values in a single band while the leaving the other bands of a spectrum undisturbed. Such artifacts may be due to different phenomena, for instance transmission and ground processing problems or single event upsets. Automated artifact detection for those phenomena is often difficult, because erroneous band values often lie well within the range of naturally occurring radiance values. We developed IDL-based software that uses edge operators to detect and label affected pixels. Using a least squares spectral matching algorithm, the distorted spectrum is compared to undisturbed spectra in the local neighborhood and the undisturbed spectrum of best fit is determined. The erroneous band value is then replaced with the corresponding undisturbed value. This method was tested on seven Landsat TM images and on artificial data. Our results show that the distorted areas are precisely detected and that the correction procedure leads to meaningful spectra. This approach may be useful to minimize the effect of single-band distortions and allows for subsequent image analysis without the need to mask out distorted areas. The software tool includes a user interface and is available online.

  Introduction

Anthropogenic land use is the main driver of terrestrial ecosystem change and results in widespread degradation and loss of ecosystem structures and functions across the globe (Foley et al. 2005). Monitoring land cover change, understanding its underlying causes, and assessing the consequences of human land use for ecosystems and biodiversity is therefore of great international concern (Gutman et al. 2004; Lambin and Geist 2006; Rindfuss et al. 2004). Remote sensing is the key technology to assess the extent and rate of land cover change (Lambin and Geist 2006; Rindfuss et al. 2004). Among the numerous earth observation satellites that are operating today, data from the Landsat Thematic Mapper (TM) 4 and 5, and the Enhanced Thematic Mapper Plus (ETM+) instruments are particularly well suited to address ecosystem dynamics. The sensors have a swath width of 185km, a spatial resolution of 30m, and six spectral bands in the visible, near- and shortwave-infrared domains (centered at 0.49, 0.56, 0.66, 0.83, 1.65, and 2.22μm, respectively). More importantly, a unique and continuous data record of Landsat images exists for the majority of the Earth’s land mass (TM-4 was operational between 1982 and 1993, TM-5 between 1984 and today, ETM+ between 1999 and today; all with a 16-day repeat cycle). Because of these properties, Landsat images continue to be the most important data source for monitoring land change at fine to medium scales (Cohen and Goward 2004; Goward and Masek 2001). Landsat data have been widely applied to assess land cover change, such as mapping tropical deforestation (Dale et al. 1993; Skole and Tucker 1993), desertification (Palmer and van Rooyen 1998), and urban growth (Seto et al. 2002; Ward et al. 2000). The research presented in this paper was carried out within the scope of several land-cover change projects focusing, for example, on Mediterranean land degradation (Hostert et al. 2003; Kuemmerle et al. 2006b) and forest mapping in Eastern Europe (Kuemmerle et al. 2006a).

Thorough pre-processing of Landsat imagery is necessary to enable multitemporal comparison and to derive accurate change maps. This usually consists of four stages: (A) correction of systematic effects such as the scan time skew, earth curvature effect, and panorama distortion, (B) radiometric calibration to convert digital numbers to at-satellite radiance (the radiant flux from a given area in the sensor’s instantaneous field of view and for a specific wavelength; measured at the sensor in W m-2 sr-1 μm-1), (C) atmospheric and topographic correction to minimize topographic influence and to attain surface reflectance, and (D) geometric correction to ortho-rectify the images. While the user has full control over stages (B), (C) and (D), the correction for systematic effects is usually carried out by the data provider (e.g. the United States Geological Survey (USGS) or Eurimage).

Additional pre-processing steps may be required if the data contains distortions or image artifacts. A multitude of phenomena are known to cause spectrally distorted pixels in Landsat TM and ETM+ images. Distortions may occur during the recording of an image or they may be introduced in the processing chain of the data vendor. Visual image analysis can compensate for these distortions to some extent, because cognitive interpretation always includes context information (e.g. spatial neighborhood information). However, image artifacts may drastically hinder automatic image analysis. This is problematic for digital change detection methods that commonly rely on the spectral comparison of a pixel at two different time stages, particularly if the images must be standardized or transformed prior to the change analysis (e.g. Tasseled Cap Transformation, NDVI calculation, etc.). Examples of such methods include image differencing, image ratioing, composite analysis, or change vector analysis (Coppin et al. 2004; Lu et al. 2004). Distorted pixels often result in the detection of pseudo-change, which can hamper or even inhibit the interpretation of the change map and the derivation of change statistics. Masking out affected areas manually is a solution, but leads to no-data areas in the change map, which is problematic for studies that address landscape pattern (e.g. forest fragmentation Kuemmerle et al. 2006a; Riitters et al. 2002). Moreover, selecting distorted pixels manually is not feasible if they are frequent or distributed over large areas. In such situations, distorted pixels need to be detected automatically and, if possible, corrected before digital change detection is carried out.

One type of image distortion common to system-corrected Landsat TM and ETM+ scenes involves individual pixels with an erroneous value for a single spectral band, while the other bands remain undisturbed. Only a minority of the erroneous band values saturates low or high and the values of the distorted pixels often are well within the range of realistic image values. The automated detection of missing pixel distortions is therefore challenging and simple threshold operations are not appropriate. A possible solution is the use of local filter operations. Erroneous band values always deviate considerably from the undisturbed values in their vicinity, thereby causing a so-called "hard edge" in brightness values in the image. Edge filters can pick up these local differences in the brightness function (Richards and Jia 2005), and by comparing edge filter images of different bands, it is possible to separate “natural” edges from edges introduced by single-band distortions.

Correcting single-band missing pixel distortions is not an easy task, because the magnitude of the distortion (i.e. the offset between the original and the erroneous band value) is unknown. However, the five undisturbed band values of a distorted pixel hold a substantial amount of spectral information that can be used to find similar, but undisturbed spectra in the local neighborhood of the distorted pixel. Erroneous band values can then be replaced using the corresponding band values of a spectrally similar pixel, thereby lessening the effect of single-band distortions, or in best case restoring the undisturbed image spectra.

The overarching goal of this study was to mitigate and correct single-band distortions in Landsat TM and ETM+ data for subsequent image processing. Our specific objective was to develop a methodology that allows for (I) the detection of distorted pixels using edge filters, and (II) finding undisturbed spectra that closely resemble the distorted spectrum to replace erroneous band values.

 Background

Missing or distorted pixels have frequently been reported to occur in the bands of the reflective domain of Landsat 4 and 5 TM instruments, and the Landsat ETM+ sensor (Helder and Ruggles 2004; Irish 2006; Saunier 2005; USGS 2006). Generally, distorted pixels can be grouped into two categories: (A) artifacts that arise from internal and external sensor anomalies during the scanning of an image and (B) artifacts introduced during the pre-processing chain of the data vendor. Distortions of type A have been discussed in the literature. Helder and Ruggles (2004) show three common radiometric artifacts, specifically the scan-correlated shift, memory effect, and coherent noise, and give advice concerning the correction of these effects. In particular the removal of striped artifact patterns known as banding or striping, which may arise from memory effects or mis-calibrated detectors, has received special attention, and various methods have been developed to correct such distortions (Gadallah et al. 2000; Helder and Ruggles 2004; Poros and Peterson 1985; Srinivasan et al. 1988). Errors of type B are not very well documented and often only in grey literature, such as data handbooks and internal handling guidelines, exists (Irish 2006; Saunier 2003, 2005).

In this study, we address Landsat TM an ETM+ image distortions that have not received much attention. These distortions can be described as pixels with obviously false or missing values in a single spectral band for a given pixel (Saunier 2003). The erroneous band value is characterized by positive or negative deviations from the actual radiance value compared to the “true” radiance values, sometimes resulting in over- or under-saturation. The Earth Observation Quality Control (EOQC) of the European Space Agency (ESA) has named the phenomenon “missing pixels effects” (Saunier 2003). Figure B-1 shows examples of such distorted pixels.

Figure B1: Three different kinds of distortions occurring in Landsat TM data. Left: only a single pixel is affected; middle: several pixels affected and random pattern of distorted pixels can be observed; right: distorted pixels are clustered and the affected show a detector pattern (all distortions occurred in a Landsat 5 TM image, acquired 4th July 1994).

A wide variety of patterns of missing pixel distortions may be observed. Generally, the phenomena may be clustered into three groups (A) single pixels with erroneous values spread out in a random pattern, which are very hard to detect visually (Figure B-1, left), (B) irregularly shaped clusters (Figure B-1, middle), and (C) a rectangular detector pattern (Figure B-1, right). Although distortions occur only in a single band for a given pixel, all optical bands of an image may be affected by the distortions, often resulting in a typical sequence of image artifacts along the scanning track of the sensor.

In most cases, only a few pixels are affected within a scene. However, image distortions may be scattered throughout the whole image and a large number of pixels may be damaged. The origin of these artifacts is not clear in all cases and an intensive discussion of sensor and pre-processing chain related image distortions is beyond the scope of this paper. However, some of the possible causes for the missing pixel phenomenon may be summarized as follows:

Sticky bits

During the recording scan, one or more bits keep their previous value instead of switching to 1 or 0 respectively. This results in an increase or decrease of the “real” digital number by a value of 2N, where N is a number between 0 and 7 (Saunier 2003).

Tape degeneration and transmission problems

Pixels or clusters of pixels may be missing due to acquisition problems or tape degradation. This problem is often reported as occurring coincidentally with other anomalies, such as data framing errors that produce scan line shifts/offsets (ESA 2003).

Ground processing problems

Pixels may become distorted during the ground recording or pre-processing of systematic effects. Cases have been reported where a scene containing missing pixel errors (e.g. dropped scans) was found to be error-free when ordered through a different data provider (G. Chander, pers. comm.). It also has been suggested that the missing pixels phenomena may be connected to old processing chains (i.e. before 1999) (Saunier 2003).

Single event upsets

A Single Event Upset (SEU) occurs when an energetic particle travels through a transistor substrate and causes electrical signals within the transistor. SEUs have been reported as often occurring in near-earth orbit when the satellite passes through the Van Allen belts (a ring of energetic charged particles around Earth, trapped by the Earth's magnetic field), especially the northern and southern auroral zones and the South Atlantic anomaly. The anomalies take the form of one or more sudden bright pixel in response to the high energy particle traveling through the transistor substrate. After the initial bright spike there are one or more dark pixels as the affected detectors recoil in bright target recovery (USGS 2000).

Although the exact sources of the missing pixels phenomena might remain unclear, it is important to mention that the effects resulting from these phenomena are similar. Thus, the method to detect and correct distorted pixels proposed in this research will work for all of these distortions, regardless of their origin.

  Methods & Materials

  Data

A total of seven Landsat TM scenes were used in this study to develop a method to detect and correct single-band missing pixels. Six images displayed such distortions, and the number of affected pixels ranged between 3,600 and 180,700. The image without distortions was used for validation purposes (see below). All images were acquired over different regions in Europe between 1984 and 1994 (Table B-1).

Table B1: Landsat TM images used in this study and the number of pixels affected by missing pixels distortions.

Path

Row

Sensor

Region

Acquisition
Date

Affected Pixels

186

026

TM 5

Eastern Europe (SE Poland)

4th July 1994

~ 108,700

186

026

TM 5

Eastern Europe (SE Poland)

7th June 1994

~ 65,400

181

035

TM 5

Southern Europe (Crete, Greece)

24th May 1986

~ 10,000

186

026

TM 5

Eastern Europe (SE Poland)

27th July 1988

0

193

023

TM 5

Central Europe (Germany)

29th April 1987

~ 42,700

188

032

TM 5

Southern Europe (Italy)

20th June 1984

~ 44,488

203

034

TM 5

Southern Europe (Portugal)

13th April 1985

~ 3,646

196

026

TM 5

Western Europe (Germany)

30th July 1984

~ 9,912

 Methods

We developed software based on the Interactive Data Language (IDL) and equipped it with a user interface (using IDL 6.0, RSI 2003). The software minimizes the effects of missing pixel distortions in Landsat images for subsequent image processing. Our methodology consisted of two stages. First, distorted pixels were detected in an automated way using edge operators. The edge filter images from different bands were then compared to separate single-band edges (i.e. edges introduced by single-band missing pixels) from multiple-band edges (i.e. natural edges). Second, we used a spectral matching procedure to find similar, but undisturbed pixels in the neighborhood of the distorted pixels. The spectrum of closest fit was used to restore the erroneous band value in the distorted pixel, thereby allowing for subsequent image processing such as digital change detection.

 Stage I: Detection of distorted pixels

Edge detection operators commonly determine the magnitude of the gradient of the continuous brightness function (x,y) to determine hard edges in an image (Richards and Jia, 2005). The magnitude of is defined by the partial spatial derivatives and (Eq. 1).

(1)

For continuous data, and are determined using Eq. 2.

, (2)

However, for discrete image data consisting of rows and columns, and are equivalent to simultaneous applications of a moving-window template in x and y direction (Richards and Jia, 2005). The Roberts filter is an easy-to-compute edge-detection operator that uses two 2x2 convolution masks (Eq. 3) to estimate and (RSI 2003).

and (3)

The Roberts operator is highly sensitive to noise, because only a very small number of pixels are used to approximate the gradient. This is usually disadvantageous, if natural edges are to be delineated, as more advanced operators with larger convolution masks are better suited for such applications. However, to detect distorted pixels, the noise sensitivity proved to be advantageous and the Roberts filter performed better than more complex operators (e.g. Canny, Prewitt, or Sobel).

Edges caused by single-band missing pixels differ from natural edges by characteristically only occurring in one image band for a certain position (i.e. a pixel). To investigate whether edges occur in multiple bands or just in a single band, band ratios were used. This concept utilizes the fact that most TM bands show a moderate to high degree of collinearity to neighboring TM bands, for instance, the three bands in the visible domain are usually highly correlated (Small 2004). To separate data artifacts from natural edges, we empirically derived band-ratio thresholds for bands that were expected to be correlated, and compared them to the band that contained the distortions (e.g. band 1 was compared with the ratio of band 1 / band 2).

While intercorrelated bands exist for the visible and shortwave-infrared domain of the Landsat TM sensors, the procedure had to be adapted for the near infrared (NIR) band (TM band 4), because some natural features only show distinct edges in this domain (e.g. edges between vegetated and non-vegetated areas). We used a moving-window Laplacian filter to identify distorted pixels in the NIR band, although relying on a single-band in the detection procedure may result in lower detection accuracies (refer to the discussion section for details). The Laplacian operator computes the second derivative to detect edges (RSI 2003). In the two-dimensional case, it is defined as the sum of the partial derivates (Eq. 4).

(4)

For discrete image data, this is equivalent to the application of a moving-window kernel (V). We used a 3 x 3 kernel (Eq. 5) to approximates the second order derivative and to detect edges in the NIR band (RSI, 2003).

(5) 

To ensure that all damaged pixels are processed in the correction procedure, the program allows for the inclusion of a user-defined buffer around the detected image distortions. A buffer of 5 pixels proved sufficient for all images used in our study.

 Stage II: Correction

Correction of distorted areas by standard operations such as low-pass or median filtering proved inappropriate because the affected areas were regularly too large. An alternative is to make use of the high degree of spatial autocorrelation that is found in Landsat images (Chica-Olmo and Abarca-Hernandez 2000; Curran and Atkinson 1998). Because real-world objects at the landscape scale are usually bigger than a single Landsat pixel, there is a high probability that a similar, but undisturbed spectrum occurs in the local neighborhood of the distorted pixels. Once such undisturbed spectra are located, they can be used to correct the distorted spectra. To find such spectra, we compared the distorted spectrum to all undisturbed spectra (i.e. all spectra not labeled in the detection procedure) that were within a defined neighborhood. For the images used in our study, a neighborhood of 100x100 pixels proved to be a good compromise between computational cost and the quality of the correction results (note that users can adjust the extent of the neighborhood in the software). As a measure of fit, we calculated the root mean squared sums (RMS) of the band-wise differences between the spectra containing the erroneous band (ui) and the band values (vi) of a given undisturbed neighborhood spectra (j) (Eq. 6, where i denotes a given undisturbed band, and k is the total number of undisturbed bands). The band containing the erroneous value was not considered in this calculation. The spectrum of closest fit was then defined as the spectra with the minimum RMS (RMS min , Eq. 7).

(6)

(7)

To correct the single-band missing pixels we then substituted the erroneous band value with the corresponding value of the RMSmin-spectrum.

Validation

Validating the effectiveness of a method to detect and correct single-band missing pixels is challenging, because the original, undisturbed values of the erroneous bands remain unknown. We therefore artificially introduced erroneous band values to nine undisturbed areas of the Landsat TM image from 4th July 1994 and tested our methodology (developed on real distortions) on these simulated distortions. Only the error-affected band values were replaced while all other band values of the original spectra remained unchanged. Altogether, 9,117 image spectra were artificially disturbed. To test our methodology, we located and corrected those pixels using our methodology and compared the corrected band values with the original, undisturbed band values. This was done by calculating the minimum, maximum, and average deviation, the standard deviation, the root mean squared error (RMSE, Eq. 6) and the confidence interval limits (p<0.01) of deviations between original and corrected band values.

To further test the usefulness of our correction procedure for subsequent change detection, we derived the normalized difference vegetation index (NDVI) and Tasseled Cap brightness, greenness, and wetness indices (Crist and Cicone 1984) for disturbed and corrected images, because both NDVI calculation and Tasseled Cap transformation are frequently carried out prior to digital change detection (Coppin et al. 2004). All calculations were carried out for a study area in the Ukrainian Carpathians (based on the image from 4th July, Table B-1). In addition, we applied three common change detection methodologies to detect forest change for our Carpathian study area using an undisturbed image from 1988 and both, the disturbed and the corrected image for 4th July 1994 (Table B-1). We used image ratioing on the raw TM images, differencing of NDVI images (Coppin et al. 2004; Lu et al. 2004), and the recently developed forest disturbance index based on Tasseled Cap transformed imagery (Healey et al. 2005). For all three change detection methods, we applied a threshold to classify the continuous change map into a binary change/no-change map. The same threshold was used for the change map derived using the disturbed image and the change map calculated using the corrected image. In the case of image ratioing, the sum of the band-wise change maps was calculated to derive a binary change/no-change map.

 Results and Discussion

The automatic detection algorithm based on edge filter images (using the Roberts operator for the visible and shortwave infrared (SWIR) bands and the Laplace operator for the NIR band) was very efficient in separating areas that were distorted by single-band missing pixels (Figure B-2). Even missing-pixels distortions that would be hard or impossible to detect visually were revealed by the detection algorithm. Such distortions include single pixels, unsaturated pixels or pixels with relatively low contrast with respect to their neighborhood, and artifacts in heterogeneous areas. Deviations between erroneous and corrected band values were substantial and varied considerably (Figure B-3).

Figure B2: Three examples of distorted areas occurring in a Landsat TM image (left), flagged pixel values after the detection procedure (middle) and the corrected image after locating similar pixels using a spectral matching operation and then substituting the erroneous band values (right). Band combinations for red, green, and blue are: 3/2/1 (top), 4/5/3 (middle), and 4/3/2 (bottom).

Validation based on artificially distorted pixels revealed very low average deviations, low standard deviations and narrow confidence intervals (Table B-2). Mean deviations were higher in the near infrared band (TM band 4) due to the high variability of vegetated areas in this spectral domain. For the imagery used in this study, the mean uncertainty level was lower than 2 digital numbers (~0.5% reflectance). The confidence intervals revealed that uncertainty connected to the mean estimation based on the sample of artificially distorted pixels was negligible. RMSE values confirmed that the vast majority of the corrected pixels displayed deviations between original and corrected band values that were much lower than the sensor’s inherent noise level (estimated by Minimum Noise Fraction images over water bodies (Table B-2). The statistical analyses also suggested that distorted band values do not over- or under-saturate in most cases.

Table B2: Statistical comparison of original and corrected spectra of artificially introduced distortions (Min / Max / Mean = minimum, maximum and average deviation of original and corrected spectra; STD = standard deviation; n = sample size; CIL / CIU = lower and upper limits of confidence intervals for p < 0.01; all values are given in digital numbers).

Band

n

Min

Max

Mean

STD

RMSE

CI L

CI U

1

1,702

0

15.00

0.35

1.06

0.60

0.29

0.42

2

1,812

0

12.00

0.19

0.62

0.44

0.16

0.23

3

2,368

0

15.00

0.43

1.19

0.65

0.36

0.49

4

559

0

15.52

1.14

2.92

1.07

0.82

1.45

5

772

0

15.78

0.78

2.18

0.88

0.57

0.98

6

1,904

0

15.52

0.46

1.40

0.68

0.38

0.55

Figure B3: Examples of spectra with erroneous band values before and after the correction. The deviation between uncorrected and corrected spectra is substantial and the correction algorithm results in more useful spectra (for details refer to text). All spectra were taken from a Landsat (5) TM image acquired 4th July 1994.

The methodology proposed in this research was highly effective in removing the missing pixel distortions for subsequent digital change detection. The artifacts were effectively removed from NDVI images and Tasseled Cap indices (Figure B-4). Even large clusters of distorted pixels were corrected, and the methodology restored textural information underlying the distorted areas, which can be important for applications that rely on continuous information on landscape heterogeneity (St-Louis et al. 2006), or for visual image analysis. Comparing the forest change maps for the three change detection methods reveals that the change maps differ substantially, because simple methods such as image differencing often results in noise in the change map. However, the change maps derived from disturbed images always label the missing pixel distortions as (pseudo-)change, thus underpinning the need to account for these distortions prior to digital change detection. After correcting the error-prone images, the missing pixels no longer appear as pseudo-change in the change maps for all three approaches (Please note that fewer missing pixels artifacts appear in the NDVI change map, because only two bands are used to derive the NDVI values, Figure B-5). The correction algorithm thus clearly mitigates and solves the single-band missing pixels problem for subsequent change analysis (Figure B-5).

Figure B4: Comparison of raw image data, NDVI, and Tasseled Cap indices for uncorrected (left) and corrected (right) images. The missing pixel distortions are not visible in the corrected images, thus allowing for better visual interpretation. All operations were carried out on a subset of a Landsat (5) TM image, acquired 4th July 1994, that displayed single-band distortions.

The methodology pursued in this research was limited to distortions that only occur in a single-band for a given pixel. Distorted pixels that display erroneous values in more than one band remained unprocessed. However, analyses of such pixels revealed that for the images used in this study, only a very small number of pixels (a maximum of 70 pixels for a full scene) were affected and hence the effect seems to be negligible. Our method can only detect distortions where the difference between the missing pixel’s value and its surrounding is large enough to be picked up by edge operators. Yet, subsequent image analysis such as digital change detection will not be hindered considerably by distortions that only slightly deviate from their neighborhood.

The Laplacian operator detected errors in the NIR band less precisely, because the Roberts operator in combination with the subsequent band ratio comparison proved to be more sensitive to subtle deviations in digital numbers. This was due to the low level of collinearity of the NIR band to other bands which in inhibited the use of band ratios. As a consequence, some single-band missing pixel distortions in the NIR band may remain undetected, although visual examination and the tests with artificial data did not suggest lower detection rates for the NIR band compared to the other bands. Using a wider buffer in the selection routine for single-band missing pixels in the NIR would be a possible approach to ensure that all distorted pixels are processed.

Figure B5: Change maps (no-change / change) for different change detection methods before (bottom) and after (top) the correction of single-band missing pixels. All analyses were carried out on the image from 7th June 1994 that had ~65,400 distorted pixels. Image ratioing (left), NDVI difference image (middle), and the disturbance index (right) were calculated for a forested region in the Ukrainian Carpathians. The forest change maps differ considerably for different change detection methods. However, the missing-pixels distortions appear as pseudo-change for all approaches when relying on uncorrected images. The correction method removes the pseudo-change from the change maps (for details refer to text).

The search area of 100x100 pixels used in this study potentially limits the correction of single-band missing pixels in cases where no matching spectrum is found in the local neighborhood. This possibly explains some outliers that existed in the statistical summary (e.g. maximum deviation of up to 15 digital numbers, Table B-2). Possible solutions include implementing a maximum threshold criterion for RMSmin (a correction is only carried out if an undisturbed spectrum with an RMSmin below a specified threshold exists), or increasing the search neighborhood (which would significantly increase the processing load).

While our method was developed to correct missing pixel distortions in Landsat TM and ETM+ images, the program is not limited to these data. Generally, distortions found in all kinds of multispectral imagery can be corrected, as long as the distortions occur in a single band and enough undistorted bands remain for the RMS calculation (to select the best undistorted spectrum). Neighborhood operations have significant potential to correct or mitigate other types of distortions, too (e.g. multiband artifacts), and future research is needed to explore these possibilities.

 Summary and Conclusions

Single-band missing pixels in Landsat TM and ETM+ datasets are relatively frequent and can tremendously hinder subsequent image analyses such as digital change detection. The software developed in this research implements a method to detect such pixels using edge operators. Using a least squares spectral matching algorithm, the distorted spectrum is compared to undisturbed spectra in the local neighborhood and the undisturbed spectrum of best fit is determined. The erroneous band value is then replaced with the corresponding value from the undisturbed spectrum.

The procedure proved to detect the affected pixels effectively. The correction yielded useful spectra and distorted areas were removed from the image. Validation of the correction algorithm using artificially distorted areas revealed that the mean deviation of original and corrected spectra was around 1 digital number and therefore well below the inherent noise level of Landsat TM imagery. Three different change detection methods carried out on uncorrected and corrected data showed that the correction prevented the missing pixels distortions from resulting in pseudo-change. The correction approach is thus a useful pre-processing step to mitigate the effect of single-band image distortions for digital change detection.

The software, programmed in IDL 6.0 (RSI 2003) and equipped with a graphical user interface, its source code, and a user guide are available free of charge to the remote sensing community via the webpage of Computers & Geosciences (www.iamg.org/
CGEditor
) and the webpage of the Geomatics Department of Humboldt-Universität zu Berlin (www.hu-geomatics.de).

Acknowledgments

The authors are grateful to E. Bowder, P. Culbert, U. C. Herzfeld, A. Janz, A. Röder, and two anonymous reviewers for helpful discussions and comments on earlier versions of this manuscript. We like to thank M. Schlerf, S. Bärisch, and the project DesertWatch, funded by the European Space Agency (ESA, project ID: ITT AO/1-4590/04/I-LG), for sharing Landsat data. Thanks are also due to G. Chander and D. A. Helder for their comments on possible sources of image distortions. We gratefully acknowledge support for A. Damm through the Nachwuchsförderungsgesetz des Landes Berlin (Young Scientists Programme of the State Berlin, NaFöG).

 References

Chica-Olmo, M., & Abarca-Hernandez, F. (2000). Computing geostatistical image texture for remotely sensed data classification. Computers & Geosciences, 26, 373-383.

Cohen, W.B., & Goward, S.N. (2004). Landsat's role in ecological applications of remote sensing. Bioscience, 54, 535-545.

Coppin, P., Jonckheere, I., Nackaerts, K., Muys, B., & Lambin, E. (2004). Digital change detection methods in ecosystem monitoring: a review. International Journal of Remote Sensing, 25, 1565-1596.

Crist, E.P., & Cicone, R.C. (1984). A physically-based transformation of Thematic Mapper data - the TM Tasseled Cap. IEEE Transactions on Geoscience and Remote Sensing, 22, 256-263.

Curran, P.J., & Atkinson, P.M. (1998). Geostatistics and remote sensing. Progress in Physical Geography, 22, 61-78.

Dale, V.H., Oneill, R.V., Pedlowski, M., & Southworth, F. (1993). Causes and effects of land-use change in central Rondonia, Brazil. Photogrammetric Engineering and Remote Sensing, 59, 997-1005.

Foley, J.A., Defries, R., Asner, G.P., Barford, C., Bonan, G., Carpenter, S.R., Chapin, F.S., Coe, M.T., Daily, G.C., Gibbs, H.K., Helkowski, J.H., Holloway, T., Howard, E.A., Kucharik, C.J., Monfreda, C., Patz, J.A., Prentice, I.C., Ramankutty, N., & Snyder, P.K. (2005). Global consequences of land use. Science, 309, 570-574.

Gadallah, F.L., Csillag, F., & Smith, E.J.M. (2000). Destriping multisensor imagery with moment matching. International Journal of Remote Sensing, 21, 2505-2511.

Goward, S.N., & Masek, J.G. (2001). Landsat - 30 years and counting. Remote Sensing of Environment, 78, 1-2.

Gutman, G., Janetos, A.C., Justice, C.O., Moran, E.F., Mustard, J.F., Rindfuss, R.R., Skole, D., Turner II, B.L., & Cochrane, M.A. (Eds.) (2004). Land Change Science. Dordrecht: Kluwer Academic Publishers.

Healey, S.P., Cohen, W.B., Yang, Z.Q., & Krankina, O.N. (2005). Comparison of Tasseled Cap-based Landsat data structures for use in forest disturbance detection. Remote Sensing of Environment, 97, 301-310.

Helder, D.L., & Ruggles, T.A. (2004). Landsat thematic mapper reflective-band radiometric artifacts. IEEE Transactions on Geoscience and Remote Sensing, 42, 2704-2716.

Hostert, P., Roder, A., & Hill, J. (2003). Coupling spectral unmixing and trend analysis for monitoring of long-term vegetation dynamics in Mediterranean rangelands. Remote Sensing of Environment, 87, 183-197.

Irish, R.R. (2006). Landsat 7 Science Data User's Handbook, Chapter 7 - Artifact Characterization, Document No. 430-15-01-003-0 [online]. Available from: [accessed 8th August 2006].

Kuemmerle, T., Roder, A., & Hill, J. (2006). Separating grassland and shrub vegetation by multidate pixel-adaptive spectral mixture analysis. International Journal of Remote Sensing, 27, 3251-3271.

Kuemmerle, T., Hostert, P., Radeloff, V.C., Perzanowski, K., & Kruhlov, I. (2007). Post-socialist forest disturbance in the Carpathian border region of Poland, Slovakia, and Ukraine. Ecological Applications, 17, 1279-1295.

Lambin, E.F., & Geist, H.J. (Eds.) (2006). Land Use and Land Cover Change. Local Processes and Global Impacts. Berlin, Heidelberg, New York: Springer Verlag.

Lu, D., Mausel, P., Brondizio, E., & Moran, E. (2004). Change detection techniques. International Journal of Remote Sensing, 25, 2365-2407.

Palmer, A.R., & van Rooyen, A.F. (1998). Detecting vegetation change in the southern Kalahari using Landsat TM data. Journal of Arid Environments, 39, 143-153.

Poros, D.J., & Peterson, C.J. (1985). Methods for Destriping Landsat Thematic Mapper Images - a Feasibility Study for an Online Destriping Process in the Thematic Mapper Image-Processing System (TIPS). Photogrammetric Engineering and Remote Sensing, 51, 1371-1378.

Richards, J.A., & Jia, X. (2005). Remote Sensing Digital Image Analysis. Berlin, Germany: Springer

Riitters, K.H., Wickham, J.D., O'Neill, R.V., Jones, K.B., Smith, E.R., Coulston, J.W., Wade, T.G., & Smith, J.H. (2002). Fragmentation of continental United States forests. Ecosystems, 5, 815-822.

Rindfuss, R.R., Walsh, S.J., Turner, B.L., Fox, J., & Mishra, V. (2004). Developing a science of land change: challenges and methodological issues. Proceedings of the National Academy of Sciences of the United States of America, 101, 13976-13981.

RSI [Research Systems, Inc.] (2003). IDL Reference Guide, IDL Version 6.0. Boulder, USA: Research Systems Inc.

Saunier, S. (2003). The Earth Observation Quality Control (EOQC), Landsat anomalies inventory - Missing image data (Slip N°3) [online]. Available from: [accessed 8th November 2006].

Saunier, S. (2005). The Earth Observation Quality Control (EOQC), Landsat frequently asked questions [online]. Available from: [accessed 8th November 2006].

Seto, K.C., Woodcock, C.E., Song, C., Huang, X., Lu, J., & Kaufmann, R.K. (2002). Monitoring land-use change in the Pearl River Delta using Landsat TM. International Journal of Remote Sensing, 23, 1985-2004.

Skole, D., & Tucker, C. (1993). Tropical deforestation and habitat fragmentation in the Amazon - satellite data from 1978 to 1988. Science, 260, 1905-1910.

Small, C. (2004). The Landsat ETM plus spectral mixing space. Remote Sensing of Environment, 93, 1-17.

Srinivasan, R., Cannon, M., & White, J. (1988). Landsat Data Destriping Using Power Spectral Filtering. Optical Engineering, 27, 939-943.

St-Louis, V., Pidgeon, A.M., Radeloff, V.C., Hawbaker, T.J., & Clayton, M.K. (2006). High-resolution image texture as a predictor of bird species richness. Remote Sensing of Environment, 105, 299-312.

USGS [United States Geological Survey] (2000). USGS Landsat project, Single event upsets (SEUs) [online]. Available from: [accessed 8th November 2006].

USGS [United States Geological Survey] (2006). USGS Landsat project, Investigations [online]. Available from: [accessed 8th November 2006].

Ward, D., Phinn, S.R., & Murray, A.T. (2000). Monitoring growth in rapidly urbanizing areas using remotely sensed data. Professional Geographer, 52, 371-386.


© Die inhaltliche Zusammenstellung und Aufmachung dieser Publikation sowie die elektronische Verarbeitung sind urheberrechtlich geschützt. Jede Verwertung, die nicht ausdrücklich vom Urheberrechtsgesetz zugelassen ist, bedarf der vorherigen Zustimmung. Das gilt insbesondere für die Vervielfältigung, die Bearbeitung und Einspeicherung und Verarbeitung in elektronische Systeme.
DiML DTD Version 4.0Zertifizierter Dokumentenserver
der Humboldt-Universität zu Berlin
HTML-Version erstellt am:
17.04.2008