Analyses performed on Landsat 5 data over the years has revealed the existence of imperfections or image artifacts caused by the instrument's electronics, dead or dying detectors, and downlink errors. As a decendant ot the Thematic Mapper, the ETM+ generates data with similar characteristics. In the past, these effects were ignored or artificially removed using cosmetic algorithms such as histogram equalization during radiometric processing. A proactive approach is in place for Landsat 7. The goal of the ground system is to remove image artifacts prior to radiometric processing. Remant artifacts, if they exist, are subsequently removed in a post-processing step using cosmetic algorithms.
The known ETM+ image artifacts are scan correlated shift, memory effect, modulation transfer function, and coherent noise. Dropped lines and inoperable detectors also exist as a result of decommutating errors and detector failure. Remnant artifacts which may exist include banding and striping. A discussion of each of these effects and characterization methodologies follows. Eliminating their presence from data products is addressed in a later chapter on Level 1 processing.
|7.2 Scan-correlated Shift|
Scan Correlated Shift (SCS) is a sudden change in bias that occur in all detectors simultaneously. The bias level switches between two states. Not all detectors are in phase, some are 180 degrees out of phase (i.e. when one detector changes from low to high, another may change from high to low). All detectors shift between two states that are constantly time varying, or slowly time varying on the order of days to months. Measurement of SCS levels is affected by another instrument artifact known as Memory Effect (ME).
Characterization of SCS is performed at three levels. First, SCS levels are obtained for each detector in each line of the scene. This must be done to all scenes that require removal of SCS. Second, the value for the SCS levels is calculated for each detector. Third, the exact location of the transition from one SCS state to another within a scan line is determined, (assuming that a continuous flow of data from the detectors is available).
|7.3 Memory Effect|
Thematic Mapper data was rife with the artifact known as ME. It is manifested in a noise pattern commonly known as banding. It can be observed as alternating lighter and darker horizontal stripes that are 16 pixels wide in data that has not been geometrically corrected. These stripes are most intense near a significant change in brightness in the horizontal (along scan) direction, such as a cloud/water boundary. Because of this, it was formerly termed 'Bright Target Saturation' or 'Bright Target Recovery.' Another artifact known as Scan Line Droop' was originally thought to be a separate phenomenon, but has since been shown to be simply another manifestation of ME. Because of its nature, ME has historically been the cause of significant error in calibration efforts since its effect on IC calibration data is scene dependent. It is present in Bands 1 through 4 of the Primary Focal Plane, and nearly absent in the Cold Focal Plane.
ME is known to be caused by circuitry contained in the pre-amplifiers immediately following the detectors in the instrument electronics. It is primarily due to a portion of a feedback circuit that contains a resistor/capacitor combination with a time constant of approximately 10 ms. This directly corresponds to time constants of approximately 1100 minor frames that have been derived from night scenes. Therefore, ME has been modeled as a simple first order linear system and only three model parameters need to be identified to characterize it (actually one of the three is simply detector bias).
|where:||g(mf) = the ME pulse response
b = detector bias
k = ME magnitude
tau = the ME time constant
Although the exact approach taken to characterize ME is dependent on data type, response of the detector to some type of pulse is measured and averaged over many scan lines. Since the response will be exponential in form, the data are manipulated by subtracting out the appropriate bias level, linearized via the natural logarithm, and linearly regressed to determine the model parameters.
|7.4 Coherent Noise|
In TM reflective band data, coherent noise (CN), manifested itself in various ways. As a consequence, it is difficult to characterize and correct for. Some CN components are locked to start of scan, but the more dominant components are not. The most persistent and dominant component is scan-free and has a varying frequency. Analysis of a swath of L5 TM night data showed the frequency of this component generally increasing as a function of time with episodic and strong jumps occurring at lamp state transitions. In addition, there is a bursting broadband component that is also not locked to start of scan. The power of this component varies strongly even within a scan, with a maximum amplitude of 1 DN. In addition, its peak frequency varies widely over a range of 0.1 inverse minor frames (imf). Consequently, this component is difficult to filter out. A common scan-locked component manifests as a spike every 16 minor frames. Its amplitude may reach up to .1-.2 DN. Another scan-locked component appears in only a few detectors. It is quite strong in one detector, having an amplitude of 0.6 DN. Interestingly, the power of this component varies significantly. Analysis of the swath of night data previously referred to, reveals its power to decrease exponentially from an amplitude of 0.6 DN to an amplitude of 0.3 DN over a timescale of about one thousand scan cycles. Finally, the power of most, if not all, CN components and the background noise (DN) correlates positively with SCS state.
The set of parameters characterizing a CN component in a detector will be the phase relative to a reference minor frame, the frequency, and the total power in the line in excess of that of the background noise. These parameters will be obtained per scan. In order to obtain the total power in excess of the background noise, the background noise must be characterized. Thus, it is necessary to characterize the continuum of the power density spectrum before identifying CN components. Analysis of the night data swath showed a strong dependence of the continuum on detector and SCS state. It also showed a dependence on lamp state, but the effect is weak. Thus, it is assumed that the continuum depends only on detector, scan direction, and SCS state.
|7.5 Dropped Lines|
Dropped lines occur in 0R data due to decommutating errors in the raw data stream ingested by LPS. During LPS processing a fill pattern is used to distinguish good data from bad data. Odd detectors are filled with zeros, while even detectors are filled with 255s. Data filling is performed on a minor frame basis - if data are missing from part of a minor frame, the entire minor frame is filled. Dropped lines can thus be entire or partial scans. Statistics on dropped lines (occurrence count, frequency) are stored in the metadata that accompanies all subintervals transferred to the LP-DAAC.
Dropped lines are characterized by examining the raw input image for the 0-255 fill pattern on a line by line basis. Known inoperable detectors are excluded from this operation. Filled minor frames are tallied by scan and for the entire image and compared to LPS counts reported in their metadata. During processing filled minor frames are flagged in the label mask which is referenced during level 1 processing.
|7.6 Inoperable Detectors|
Image artifacts caused by dead or dying detectors require characterization. In addition to the obvious case of a "dead detector," one which provides no change in output DN for changes in input radiance it is also important to determine when a detector channel has fallen out of acceptable performance limits. A starting point for such a determination is a test to see if each detector meets the performance criteria established in the Landsat-7 System Specification, Para. 188.8.131.52.3.1. This specification provides a detector to be classified as degraded, if its SNR or dynamic range are below the specification levels.
This is a distributed algorithm that evaluates the outputs of several of the characterizations. One piece takes the indications of dead detectors from the output of perform IC to determine detector "aliveness". A second piece takes the outputs of random noise characterization to determine "in-spec" or out of spec behavior of the detector noise levels. This latter piece has two portions: one of which can operate on any scene using the shutter data, the other operates only on the output of FASC scenes. The final piece of the algorithm takes the output of the current gain selection and the currently indicated saturation bins, to assess the dynamic range of the channels. If any portion of the algorithm detects a change in performance (e.g. a new dead detector, noise level below spec or dynamic range below specs) flags will be set for an analyst to examine the results and potentially update the parameter file.
Banding or "scan-to-scan striping", is a sometimes visible noise pattern caused by memory effect. After scanning past a bright target such as clouds or snow, detector response is reduced due to memory effect. Thus if the region past the bright target is uniform, data values obtained from the sensor will be slightly lower than corresponding values obtained on the following scan (since the following scan is in the opposite direction and therefore, has yet to encounter the bright target.) The result is that scans in one direction will be noticeably darker than adjacent scans in the opposite direction. The banding pattern is very small in intensity, typically on the order of 1 to 2 counts.
Banding characterizations are performed after artifact removal and radiometric correction. The operation is cosmetic in nature as it recognizes and removes striping patterns left over from an ineffective ME correction process.
The approach employs an algorithm that applies a filter optimized for detection of the banding pattern. Because of the small amplitude inherent to banding, the filter is adaptive so that it only operates on those portions of the image where banding is detectable (i.e. in homogeneous or "smooth" image regions.) The filtering operation produces an output image where banding has been removed, as well as a difference image that gives an indication of where banding was detected, and it's amplitude. An overall figure of merit is also calculated.
Striping is a line-to-line artifact phenomenon that appears in individual bands of radiometrically corrected data. Its source can be traced to individual detectors that are miscalibrated with respect to one another. The application of the calibration coefficients to the ETM+ data, i.e. the generation of the level 1R data, is intended to remove the detector to detector variations in gain and offset, effectively de-striping the data. As detector to detector variations are already explicitly taken into account through the generation of relative gains and bias from histograms, and these are included in the process of generating the applied gains and biases, the striping characterization and correction should not be required in routine processing.
Nonetheless, a post processing characteriztion and removal process is in place should a cosmetic fix become necessary. This is achieved by linearly adjusting the 1R data to match the means and standard deviations of each detector to a reference detector, or to the average of the detectors. This actual processing algorithm applies the relative gains and biases calculated by the histogram analysis performed on every image just prior to radiometric correction.