Mosaicking of global planetary image datasets: 1. Techniques and data processing for Thermal Emission Imaging System (THEMIS) multi-spectral data



[1] The mosaicking of global planetary data sets allows for the examination of local, regional, and global scale processes on all planetary bodies. Processing techniques that allow us and other users to crate mosaics of tens of thousands of images are documented along with the associated errors introduced by each image-processing algorithm. These techniques (e.g., non-uniformity correction, running contrast stretches, line and row correlated noise removal, and random noise removal) were originally developed for the 2001 Mars Odyssey Thermal Emission Imaging System (THEMIS) infrared multispectral imager data but can be adapted and applied to other data sets by the alteration of input parameters. The techniques for mosaicking planetary image data sets (e.g., image registration, blending, and normalization) are also presented along with the generation of qualitative and quantitative products. These techniques are then applied to generate THEMIS daytime and nighttime infrared, Viking, Context Imager (CTX), and Mars Orbiter Camera (MOC) visible mosaics using a variety of input and output types at a variety of scales. By creating mosaics of the same area using different data sets such as those that illustrate compositional diversity, thermophysical properties, or small-scale morphology, it is possible to view the surface of the planet and geologic problems through many different perspectives. In addition to the techniques used to create large-scale seamless mosaics, we also present the THEMIS daytime and nighttime relative temperature global mosaics, which are the highest resolution (100m/pixel) global scale data sets available for Mars to date.

1. Introduction

[2] Images of other planetary bodies in our solar system are some of the most widely utilized data products available to the planetary science community. Mars has a wealth of these data, which have been acquired from the beginning of NASA's exploration of the solar system to the present-day. Imaging cameras and spectrometers such as the Viking Orbiter Visual Imaging Subsystems (VIS) [Klaasen et al., 1977], the Mars Orbiter Camera (MOC) [Malin et al., 1998] wide-angle and narrow angle instruments, the Thermal Emission Imaging Systems (THEMIS) [Christensen et al., 2003, 2004] visual and infrared imagers, the High-Resolution Stereo Camera (HRSC) [Neukum et al., 2004; Jaumann et al., 2007] visible imager, and the Mars Reconnaissance Orbiter's High Resolution Imaging Science Experiment (HiRISE) [McEwen et al., 2007], Context Imager (CTX) [Malin et al., 2007], and the Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) [Murchie et al., 2007] have all provided new and unique views of the planet that have revolutionized the manner and detail in which Mars is studied. While many of these instruments do not provide global scale coverage in a single image (in fact many instruments only provide meter to kilometer scale snapshots of the surface), the combination of hundreds to tens of thousands of these individual observations provides a remarkably powerful tool for scientific investigations.

[3] The construction and use of global data sets has vastly increased the present understanding of many aspects of Mars, including its crustal and regolith composition, thermophysical character, magnetic field anomalies, active processes, climate, and geologic history. Here we will document the processes that are used to create large-scale regional and global mosaics of thousands of individual Thermal Emission Imaging System (THEMIS) images and demonstrate the versatility of these tools by applying them to other planetary image data sets including Viking VIS, MOC, and CTX. The companion paper to this publication, Edwards et al. [2011], hereinafter referred to as Paper 2, uses the THEMIS daytime and nighttime infrared mosaics as reconnaissance tools and presents unique scientific results made possible by the creation of the mosaic.

[4] THEMIS provides excellent means to address many local and regional scale geologic problems, with its high spatial sampling and its ability to discriminate between geologic materials of different compositions. These compositional data have allowed for the investigation of a variety of areas in detail, including the Nili Fossae [Hamilton and Christensen, 2005], Ganges and Eos Chasmata [Edwards et al., 2008], Ares Vallis [Rogers et al., 2005], Mare Serpentis [Rogers et al., 2009], and a variety of other locations. The majority of these studies utilized THEMIS data in conjunction with Thermal Emission Spectrometer (TES) [Christensen et al., 2000, 2001], CRISM [Murchie et al., 2007; Pelkey et al., 2007] and Observatoire pour la Minéralogie, l'Eau, les Glaces, et l'Activité (OMEGA) [Bibring et al., 2005; Mustard et al., 2005] data to better constrain the compositions highlighted by THEMIS data. These compositional data have also been used in conjunction with THEMIS thermophysical data [Fergason et al., 2006; Bandfield, 2008; Bandfield and Rogers, 2008; Edwards et al., 2008] to constrain the nature of the material (e.g., unconsolidated sand or dust). Additionally, these data have allowed for the investigation of in situ bedrock or rocky materials [Edwards et al., 2009] and primarily aeolian deposits in Arabia Terra [Fergason and Christensen, 2008], where likely the most primitive unmodified crustal and the most homogenized materials, respectively, are observed on the planet.

[5] In this paper, we present the highest resolution global data sets of Mars to date with a spatial sampling of 100m/pixel or ∼592 pixels per degree (ppd) at the equator. Both nighttime temperature and daytime temperature qualitative global data sets have been produced at this scale. While these data do not provide quantitative thermophysical or compositional information about the surface, they are extremely useful as reconnaissance tools. This is especially true for areas with materials that exhibit a range of thermophysical characteristics, which can be readily identified. In addition, the daytime data can be used for morphologic studies, crater counting, and as indicators of thermophysical properties or compositional variations. The construction of qualitative data products is the initial step required for subsequent work to construct global thermal inertia, multiband decorrelation stretch, 10-band emissivity, and surface kinetic temperature mosaics at the same 100m/pixel scale.

2. Data Processing Tools

[6] The construction of these data sets necessitates the use of advanced image processing techniques that were originally developed for use with THEMIS image data products but can be adapted to a variety of other data sets through the alteration of input parameters. These techniques are used to create well-calibrated high quality quantitative data from which additional products can be derived, as well as large-scale qualitative image products, which are equally useful and provide regional morphologic and relative thermophysical context. Often, the size of the region of interest will require the concatenation of many hundreds to thousands of images (or tens of thousands as is the case for the THEMIS daytime and nighttime global mosaics).

[7] Here we present the generic techniques and tools which can be used to create mosaics from any images projected by the Integrated Software for Imagers and Spectrometers (ISIS; package provided by the United States Geological Survey (USGS) [e.g., Torson and Becker, 1997; Anderson et al., 2004]. ISIS provides many useful processing tools (projection and other geometric processing of images with high precision, based on rigorous camera models, precise radiometric calibration for many instruments, and a large but certainly not complete set of other cartographic and image processing functions such as map transformations and spatial filtering), but is not provided to the end user as a development platform for new algorithms. It supports predetermined processing steps used to manipulate data in a standard and established manner.

[8] Other than map projecting data to a planet's surface, all data processing and mosaicking was completed using an open source software package entitled DaVinci, which is maintained by the Mars Space Flight Facility at Arizona State University ( in partnership with other institutions. DaVinci is a generic array-processing tool that allows for symbolic and mathematical manipulation of hyper-spectral planetary image data and provides support for importing and exporting current ISIS data formats. DaVinci allows the end user to develop image-processing algorithms with an interactive scripting interface. Its plotting and image display capabilities let the user visualize the effect of data processing in real-time. Processing algorithms developed in DaVinci can be easily integrated with ISIS to provide a flexible compliment to the established ISIS routines.

[9] ISIS currently supports Linux and MacOS X installations and requires up to 70 Gigabytes of storage for the application and mission data. DaVinci supports Linux, MacOS X, and Windows installations and only requires 100 Megabytes of storage for the entire application. Both applications have detailed documentation, tutorials, and additional information that can be found at their respective websites.

[10] A combination of ISIS and DaVinci was used to create mosaics for several different data sets including the Viking VIS [e.g., Klaasen et al., 1977], MOC wide-angle camera [e.g., Malin et al., 1998], CTX [e.g., Malin et al., 2007] and THEMIS infrared and visible imagers [Christensen et al., 2003, 2004]. In addition, we demonstrate the efficacy of the aforementioned techniques and software to create regional multiband mosaics using decorrelation stretch [Gillespie et al., 1986] THEMIS images that vividly illustrate the compositional diversity of the Martian crust.

3. THEMIS Specific Data Processing

3.1. Data Set/Instrument Description

[11] The THEMIS instrument includes two separate cameras: a thermal infrared pushbroom line scanner consisting of a 9 spectral channel, 320 by 240 element uncooled microbolometer array that covers a wavelength range of 6.7 to 14.8 μm and a visible 1024 by 1024 framelet imager split into 5 spectral channels in ∼50nm bands centered from 0.42 to 0.86μm [Christensen et al., 2004]. The infrared imager has 100 m/pixel spatial sampling and the visible imager has ∼18m/pixel spatial sampling from the ∼420 km altitude orbit of the 2001 Mars Odyssey spacecraft. The infrared imager acquires extremely long (typically >6000 lines, covering >600km), narrow (320 samples, covering only 32km) images. Several issues commonly exist with line scanner instruments like THEMIS including: 1) variations in detector readout voltage that can cause enhanced line-to-line noise, 2) the delay between the acquisition of each band that necessitates the use of map projection software to align each band to the surface before the multispectral data are utilized, and 3) calibration errors as these instruments will often acquire data for long periods of time covering large distances on the surface.

[12] Candidate images are first selected using parameters in a database (available through a DaVinci Simple Query Language (SQL) interface or and are chosen according to the desired use of the mosaic. For example, it is possible to select images based on surface temperature, atmospheric opacities, thermal inertia, TES albedo, incidence angle, solar longitude, and a variety of other acquisition and observational parameters [e.g., Christensen et al., 2007]. Additional quality control constraints are useful to eliminate noisy, poorly calibrated data. Data with the following parameters are preferred: 1) the time difference between the end of the image and the acquisition of the calibration image should be limited to <150 s to minimize the potential changes in instrumental conditions, 2) images with a small number of saturated or under-saturated pixels, 3) images with low percentages of data dropouts, and 4) images that do not have enhanced line-to-line noise.

[13] Prior to mosaicking, visual examination of every image is often included as an additional step. Data that are often excluded include images containing elevated line-to-line or white noise and excessive repeated dropouts, which render the blending of several images difficult, and images collected during periods of high atmospheric dust where the overall contrast of the image is reduced. The highest quality images correspond to data with the warmest surfaces and lowest atmospheric opacity, although for qualitative, single-band mosaics, data with lower surface kinetic or brightness temperatures are often acceptable. Other instruments (both line scanner and frame imagers) have a similar set of numeric and visual selection criteria that will be applied to arrive at a high-quality image set for mosaicking.

3.2. THEMIS Global Mosaic Data Selection

[14] Perhaps the most important aspect of the entire mosaicking process is the selection of high quality, well-calibrated data. During the THEMIS mapping mission, images were both systematically and priority targeted, potentially covering the same area on the planet many times over. Data were initially selected following the guidelines in section 3.1; however, when an area was imaged multiple times, the data with the highest visual quality were selected and other lower quality data were removed, reducing the probability of blending poorly registered data.

[15] Equatorial and midlatitude data from −60°N to 60°N were limited by a solar incidence angle <85° for daytime and >95° for nighttime, ensuring that the sun was at least 5° above or below the horizon respectively. Additionally, only full-resolution (unsummed) data were used. Images where the time between the end of image acquisition and the collection of calibration data exceeds 150 s were omitted as a quality control measure, helping to ensure the usage of well-calibrated data. Longer times between the data collection and calibration image acquisition translate into larger uncertainties due to focal plane temperature drift. Images with any oversaturated or undersaturated pixels identified in the initial image calibration (these pixels have DN values at the minimum or maximum of the 12-bit THEMIS instrument and are logged as the fraction of saturated pixels to the total number of pixels in the image) were also rejected as these pixels can introduce undesirable image artifacts, negatively influencing the overall image quality and visual appearance in the final mosaic product.

[16] Only the daytime mosaic was constructed for latitudes poleward of 60° as nighttime surface temperatures are exceedingly low (e.g., <160K) and instrument signal-to-noise is low at these temperatures. As the THEMIS global mosaics are designed to maximize coverage and not minimize seasonal effects, no restriction on season was placed on daytime images poleward of 60° latitude. Seasonal effects in daytime data could cause significant temperature variations between images; however, these variations are most obvious in nighttime temperatures. Constructing a daytime mosaic that incorporates all seasons is a reasonable task for the mid- to upper-latitudes [e.g., Bandfield, 2007; Bandfield and Feldman, 2008]. Near the Martian poles, the presence or absence of seasonal ice and clouds is another issue that must be addressed [e.g., Kieffer et al., 2000; Kieffer and Titus, 2001; Piqueux et al., 2003, 2008]. Polar features are highly variable and can change on a daily basis. No attempt, other than removing images with large areas of seasonal CO2 ice, has been made to eliminate seasonal variations; therefore polar regions (as with the rest of the global mosaics) are a combination of data from a variety of seasons and conditions. Without the use of the image processing techniques discussed in this paper, we would be unable to create seamless global-scale mosaics composed of tens of thousands of images, which are acquired at different seasons and times of day.

3.3. Standard THEMIS Processing

[17] THEMIS data are provided as Reduced Data Records (RDR) through the Planetary Data System (PDS,, which are radiometrically calibrated radiance data derived from the THEMIS Experimental Data Record (EDR) [Christensen, 2002] and THEMIS calibration report ( [Christensen et al., 2004]. Through the completion of the mosaicking process, images remain consistent to the PDS RDR calibrated radiance form, retaining all spectral bands, processing history, and data precision, as it provides the most flexibility to convert the data to other products (e.g., emissivity, stretched single- and multiband images, surface temperature etc.) and quantitative data retained until the very last step where the mosaic is scaled for viewing. Some processes require pixel-to-pixel registration (e.g., band-dependent line- and row-correlated noise removal and random noise removal) provided by the THEMIS camera model using the ISIS software and all data products after an ISIS processing step are consistent with the ISIS format, again retaining all spectral bands, processing history, and data precision.

[18] In order to display the data for viewing, several different methods must be taken to scale the data and maximize the desired features. These techniques range from single band histogram and linear stretches to multiple band principal component analysis and decorrelation stretches [e.g., Gillespie et al., 1986]. Decorrelation stretches are the predominant technique that is utilized in this paper. In THEMIS data, decorrelation stretches maximize the compositional/spectral variation (which is related to the second and third principal components), while retaining much of the morphologic/temperature context (which is captured primarily in the first principal component). This type of stretch was chosen because it is extremely good at highlighting second and third order spectral variations, which is the main focus of the THEMIS data processing presented in this paper. Additionally, decorrelation stretches are used throughout this paper to provide a consistent means of viewing the effects of the different algorithms.

[19] A flowchart illustrating the order, input, output, and changes to the data are shown in Figure 1. This figure also includes corresponding section numbers and information about the mosaicking procedure described below, as well as specific function names for both DaVinci and ISIS. Unless otherwise stated, all of the following algorithms and processes have been completed using the DaVinci software package.

Figure 1.

This set of flowcharts details the processing steps that are taken to process both THEMIS data and qualitative/quantitative mosaics. For each processing step a reference to the corresponding section in the text is provided on the arrow, as well as a short description of the step, the software package used (DaVinci or ISIS) and the function name that completes the task adjacent to the image set. (a) THEMIS standard processing flowchart illustrating the typical steps used to process THEMIS data into the form commonly used to mosaic these data. THEMIS data are ingested at the RDR state from the Planetary Data System (PDS) archive and then follow the arrows in the diagram through the various processing stages. Each of these stages (except for the map projection) is optional and may be excluded depending on the final processing state required. Three images are shown where the leftmost is the input image, the center is the output data from the algorithm, and the rightmost image is the difference between the two. Each set of three images has been stretched using a decorrelation stretch with the same statistics so a direct comparison can be made between the image set. (b) The standard mosaic processing flowchart illustrates the two-path (qualitative and quantitative) nature of the software. If the desired end result is qualitative then processing steps that disrupt the calibration of the data are taken to provide a visually appealing mosaic. However, if the desired end result is quantitative, then no normalization steps are taken, preserving the nature of the input data. Three sample data types (Thermal Inertia, Emissivity, and Surface Temperature) are shown along with the associated papers [Christensen et al., 2003; Bandfield et al., 2004; Fergason et al., 2006] that describe the techniques required to process the input data to these stages. Any number of additional data types may be created here and will follow the same path as the quantitative products.

Figure 1.


[20] The RDR data are first corrected for time-dependent focal plane temperature variations in the detector array which causes a ± 1DN offset at scales of ∼50–200 line in the image [Bandfield et al., 2004]. Additionally, the image is corrected for lower frequency temperature drift (±10 DN on the dayside and ±5 DN on the night side), which is related to the constantly changing position of the instrument relative to the Sun and the temperature of the observed target. These offsets are determined using the THEMIS atmospheric channel (band 10), with the assumption that the atmosphere does not have high frequency temperature variations [Bandfield et al., 2004].

3.4. Advanced THEMIS Data Processing

3.4.1. Temperature Variation Across the Calibration Flag Correction

[21] Following the removal of time-dependent focal plane temperature variations and the low frequency temperature drift, a temperature variation across the calibration flag is removed. This effect creates an apparent cross-track slope in the temperature image. The amplitude of slope is typically small (<1K), but can cause difficulty in making seamless mosaics; the temperature gradient across the image results in artificially bright regions in one image being abutted to artificially dark regions in its neighbor. As is nearly always the case, the temperature variation across the calibration flag is consistent in its slope (e.g., always negative, or left to right in an image). A calibration flag is the primary means by which THEMIS data are converted from raw data numbers to calibrated radiance. The calibration flag is a known temperature that is imaged by the microbolometer array after every image acquisition and provides a single point to relate back to the THEMIS instrument response function lookup table. The temperature variation across the calibration flag is variable, but not independent for each band, as the bands are not acquired simultaneously while observing the same position on the planet (or in this case calibration flag). The variation is also visible in decorrelation stretch images where it shows up as a color variation (e.g., red to blue tones, green to blue tones) from the left side of the image to the right side of the image (Figures 2a and 2c) depending on the selection of bands. Figure 3 illustrates the absolute temperature difference for two bands before and after the calibration flag temperature variation removal algorithm is applied. The observed difference in slope between bands 4 and 9 is the cause of the color variation observed in the decorrelation stretch images in Figure 2.

Figure 2.

These images are decorrelation stretches of the same THEMIS image (I26686040, bands 9, 6, and 4). (a) The original THEMIS image before the tilt removal algorithm has been applied. Tilt is manifested as a red to blue to green shift (from left to right) in this case. (b) The same image that has undergone the tilt removal process, which has removed the left to right color variation. (c) The difference of the radiance between the images from Figures 2a and 2b where the data have been decorrelation stretched using the same stretch as the image in Figure 2b. This illustrates what has been removed from the uncorrected data to produce the corrected data.

Figure 3.

Row averaged temperature plotted for a relatively long THEMIS image (I26686040, > 10000 lines) illustrates the effect of tilt. The solid lines are the original data and the dashed lines have had the tilt removal algorithm applied. At most the temperature difference between the solid and dashed lines is ∼±1K and the tilt is interpreted to be the slope from left to right (in this specific instance). The dashed line has been normalized for this slope, while retaining the spatial information. The offset between bands 4 and 9 is artificial, for clarity; however, the difference in slopes between the bands is real.

[22] In addition to varying across the image, the temperature variation is not constant throughout the duration of the image, implying that some residual time-dependent parameter is not being completely removed and may be related to another instrument component. In order to correct for these variations, we average all the columns of the image in 4000 line segments and convolve this average with a filter comparable to the width of the image (e.g., 320 samples), which is then subtracted with the overall average value of the section added back. We allow for 2000 lines of overlap between sections to ensure that the time-dependent parameter is removed and then perform a linear ramp over the overlapping region to ensure that a smooth transition from section to section is achieved. This processes is called a “running” or windowed process, as it operates upon independent but multiply redundant sections of the image simultaneously and linearly combines overlapping pieces with fractions of the data ranging from 0 to 1 (depending on their position) and relative contributions to create the final output. In this case, the algorithm is only applied in one direction (along track), but for other processes the running algorithm may be applied in two or three dimensions with differing amounts of overlap.

3.4.2. Band-Independent and Band-Dependent Row and Line Correlated Noise Removal

[23] The corrected data are then map projected using the ISIS software, which utilizes a camera model to accurately describe the mapping of detector pixels to the surface of Mars. Once the bands are spatially registered, a form of row- and column-correlated band-independent noise, also known as “plaid,” is apparent in decorrelation stretch and single band nighttime images. This form of row- and column-correlated noise is primarily a spectral contribution (e.g., varying for each band) and is identified by abnormally high or low row- or column-average values in one spectral band but not any of the other spectral bands. This effect was first noted by Bandfield et al. [2004] and the correction algorithm and cause of the plaid is described in detail by K. J. Nowicki and P. R. Christensen (Removal of line- and row-correlated noise in THEMIS multi-spectral infrared data, submitted to Mars, 2011). The row-correlated component of plaid is residual electronic noise that is mapped into the RDR during the read-out of the array and consequent transfer to the spacecraft memory buffer. The column-correlated component of the plaid is caused by variations across the detector that are not adequately removed by the instrument response function. These artifacts are typically at the level of ± 1DN and are only evident in decorrelation stretch images or other extreme processing approaches such as image ratioing or principal components analysis that emphasize small differences between spectral bands. However, similar artifacts may become increasingly evident with lower signal to noise (SNR) data, as is the case with a significant portion of nighttime infrared images and in instruments with a lower SNR than THEMIS.

3.4.3. Random Noise Removal

[24] In decorrelation stretch images, pixel-to-pixel white noise can be significant, obfuscating surface features and the interpretation of false color images. Spectral differences between adjacent pixels are maximized by the decorrelation stretch and can often account for a large fraction of the spectral variability in a scene (Figures 4a and 4c). We use a white noise removal algorithm to reduce the random pixel-to-pixel noise and accentuate the underlying surface spectral variation in the component images of mosaics (Figures 4b and 4d).

Figure 4.

(a) The original THEMIS data (I25647004) decorrelation stretched with THEMIS bands 8, 7, and 5 illustrates the white noise that is the focus of this example. The white noise in this data is characterized by multicolored speckles distributed randomly throughout the image. (b) The same image data with the white noise removal algorithm applied exhibits much less of the speckled texture that was observed in the companion image. Additionally, features that are not visible in the original data are easily observed in this new image, as the amount of random or white noise has been significantly reduced. (c) A subsection of the image in Figure 4a illustrating the level of the white noise. (d) A subsection of the image in Figure 4b illustrating the effect of the noise removal algorithm.

[25] In this algorithm, each band of THEMIS radiance is converted into a percentage of the total signal. The total signal is the integral of all of the radiances in all of the surface sensitive bands. To convert each band into a percentage of the total signal, we divide the radiance of each band by the total signal, yielding a fractional contribution image with values ranging from 0 to 1. Each band of the fractional contribution image is then typically convolved with a 7 × 7 pixel boxcar filter and then multiplied by the total signal image to regain original radiance units. The assumption behind the algorithm is that no two adjacent pixels in any band should contribute a significantly large difference in the percentage of the total signal. This assumption breaks down at bright and dark (or hot and cold) boundaries where artificially high or low radiance values are imparted in the data by the convolution. The difference between the original radiance and the newly filtered radiance is largely composed of random noise, plus the artificially elevated or lowered boundary data. The difference image is then converted to a principle component image, of which the first principle component is the correlated boundary data. The first principle component is removed and the image is converted back into radiance units, resulting in an image with only randomly distributed noise. The randomly distributed noise is subtracted from the original radiance image to obtain the noise-free radiance data. To obtain high correlation values in the principle component image, this entire operation is calculated for every 1001 by 1001 pixel segment of an image in the same manner as the running stretches described in section 4.1.

[26] Figure 5 illustrates the absolute radiance difference between the corrected and uncorrected data for several THEMIS bands. The maximum correction in this case corresponds to ∼1 × 10−5 W cm−2μm−1 sr−1 with an average correction of −1.6 × 10−12 ± 1.4 × 10−6 W cm−2μm−1 sr−1. Outside the case of decorrelation stretches, the alterations to the data from this algorithm are undetectable and functionally insignificant. For a more detailed discussion of this algorithm and the errors associated with its use, see K. J. Nowicki et al. (Removal of white noise from multi-spectral THEMIS data, manuscript in preparation, 2011).

Figure 5.

THEMIS image (I25647004) band 2 (blue), 5 (green), and 9 (red) uncorrected radiance plotted versus the difference between white noise corrected and uncorrected radiance for the corresponding bands. In this case, every pixel in a THEMIS image with the white noise removal algorithm applied is subtracted from the corresponding pixel in the original data. This allows for a direct comparison of the effectiveness and scale of the correction. If no correction was applied, the data would follow the horizontal axis. The spread in the data is roughly the same for each band with the maximum correction corresponding to ∼±1 × 10−5 W cm−2μm−1 sr−1 and an average correction of −1.6 × 10−12 ± 1.4 × 10−6 W cm−2μm−1 sr−1.

4. Mosaicking Algorithms and Procedure

[27] The DaVinci software package provides the ability to load ISIS data cubes (among a variety of other file formats) into a manageable format and extract the required information for mosaicking, including mapping parameters such as the coordinate system, the line and sample projection offset, resolution/map scale, image width and height, etc. as well as quantitative spectral and temperature data. If a single band mosaic is desired, then a single band is extracted from the ISIS data cube. At the wavelengths covered by THEMIS band 9, surface emissivity is high (∼0.99), atmospheric opacity is low, and the surface-atmosphere temperature contrast is high, making this the most common THEMIS band utilized in the mosaicking process as it has the highest signal-to-noise ratio for both daytime and nighttime data.

[28] Once the data are ingested and the bounds for the mosaic have been set, either by the maximum extent of the images used or a user defined latitude/longitude range, the appropriate line projection offset and sample projection offset are used to place the data into the context of the mosaic region for the given projection. Figure 1b illustrates the typical order and results of the mosaicking algorithms described below. Two paths are shown, where one is the path taken for a qualitative product while the other path is taken for a quantitative end result. The algorithms used to create the quantitative data products are not discussed here in detail as they vary significantly between the data being derived and the science objectives [e.g., Christensen et al., 2003; Bandfield et al., 2004; Fergason et al., 2006].

4.1. Running Stretches

[29] After the initial image position in the mosaic region has been determined, several additional processing steps occur to ensure that the final mosaic is seamless. As THEMIS infrared images are commonly 10 to 20 times longer than they are wide and the 2001 Mars Odyssey is in polar orbit, images cover large ranges of latitude (often >15°), where surface temperatures may vary significantly. Several processes are performed on the data to normalize for the temperature difference.

[30] First, a large (often > 1001 by 1001 pixels) high pass filter can be applied to the data to remove the low frequency change in surface temperature, caused by the differences in acquisition latitude. Additionally, similar to the temperature variation across the calibration flag correction algorithm, a running histogram or decorrelation stretch can also be preformed to maximize the local variation in the scene. For THEMIS, these stretches are typically done on ∼1000 by 1000 line sections in two dimensions with 50% overlap of each section as this provides the best image-to-image normalization for long images. An example of the small-scale detail that is enhanced with this method is shown in Figure 6 where running histogram stretch was applied to the data. In cases where the final product is not intended to preserve a physical quantity such as temperature, the data are stretched prior to being inserted into the final mosaic region, yielding better contrast-matched mosaics.

Figure 6.

Two different stretches of the same THEMIS image (106568023, band 9 radiance) presented side by side, illustrating the advantage of a running histogram stretch. An image of Valles Marineris was chosen to highlight the strengths of the running histogram stretch. Similar advantages are observed with running decorrelation stretches. (a) A standard histogram stretch of a THEMIS image, where some saturated and undersaturated locations are present in the canyon and craters walls. Additionally, the plains have low contrast, as the statistics for the entire scene are being dominated by warm and cold locations in the canyon walls. (b) A running histogram stretch of the same image. In this case, the canyon walls are less saturated than in the standard case. However, the most significant differences are observed in the plains, where details that were not visible previously are now easily observed. However, the data values associated with this stretch are relative to one another over the window size of the algorithm, whereas in the standard histogram stretch case all the values in the entire image are relative to each other. For example, the stretched values of pixels containing the same calibrated radiance values may not be mapped to the same stretched value, or pixels with significantly different radiance values may end up with similar stretched values.

4.2. Automated Image Registration

[31] An automated image registration algorithm can applied to correct for potential location inaccuracies due to uncertainties in the timing of the start of the image or spacecraft pointing. In THEMIS data this uncertainty results in a ±2 pixel (200 m) error in absolute position on the surface of Mars in the along track direction. There may also be absolute errors in cross-track position, which for THEMIS are typically <1 pixel (100 m) that is related to the inaccuracies in the derived spacecraft position and pointing. In the case of THEMIS data, only one offset is stored for each image and individual bands of an image are not shifted independently, as the THEMIS camera model in ISIS aligns the individual bands to <0.1 pixel.

[32] Spacecraft pointing errors are typically small for THEMIS infrared and visible data; however, for older data sets where the position and pointing of the spacecraft was not as well known, (e.g., Viking), the offsets due to pointing errors can be significant, often greater than several kilometers (e.g., 100s of pixels). For this algorithm to be applied, a reference image or image location must be chosen, from which to base all subsequent positioning of images. For consistency, the images are typically sorted from west to east, but may also be sorted by maximum overlap. The auto-registration algorithm is a two-dimensional difference minimization technique, where a random sampling of pixels in the original overlapping image pairs are taken and a search radius of R number of pixels are compared to the sum of the difference of the values of the two images. The process is repeated until the algorithm arrives at the local minimum value of the solution space. Implementation of a random sampling technique and the auto registration algorithm is applied twice to ensure that the minimum reached in the first solution is in fact the true minimum and not a local minimum. It is possible to perform an exhaustive search of the overlapping region, where every pixel is compared to every other pixel in the search radius, but these operations are computationally costly, especially when this process must be completed for each overlapping image (up to several thousand times). Experience has shown that the random technique produces a comparable result to an exhaustive search.

4.3. Image to Image Blending

[33] As a final step the data are blended using a two dimensional linear combination ramp of the overlapping regions to insert the new image into the mosaic and is stretched as the last step just prior to insertion into the final region. This blending algorithm does not blend all images at once; rather each subsequent image is blended into the final output image and the resultant image can be the average of many images. However, this algorithm will favor the most recently added image, as the overlapping sections are only a weighted average of what exists in the destination image (which could be many previously averaged images) and the newly inserted image. In summary, for qualitative data products, the best results are typically achieved when a large high pass filter is applied to remove the low frequency change in surface temperature, followed by the stretching of individual images (either by a running or standard stretch type), and then a multidirectional blend of adjacent images, which are inserted into the final mosaic scene.

4.4. Quantitative Mosaics

[34] Data normalization steps such as nonlinear stretches and running stretches (section 4.1), high-pass filters (section 4.1), and random noise removal (section 3.4.3) remove the quantitative aspect of data and should be excluded in the creation of quantitative mosaics (Figure 1b). However, other empirical data corrections such as the removal of the temperature variation across the calibration flag (section 3.4.1) and the removal of band-dependent and band-independent line- and row-correlated noise (section 3.4.2) result in data that have values that more closely represent actual conditions and are less effected by instrumental artifacts, making them desirable corrections to produce accurate and well calibrated data. When creating a quantitative mosaic, the use of additional algorithms to compute the desired value (e.g., thermal inertia [Fergason et al., 2006] or emissivity [Christensen et al., 2003]) commonly have the effect of normalizing data in a quantitative manner. When these types of processes can be applied, they are preferred over the less well-characterized data normalization algorithms presented in this work. Figure 1b illustrates the two-path approach (qualitative on the left and quantitative on the right) that is taken by the mosaic software where three sample data types (emissivity [Christensen et al., 2003], surface temperature [Bandfield et al., 2004], and thermal inertia [Fergason et al., 2006]) are shown. While these three data examples are specific to THEMIS, it is possible to derive any desired data products such as mineral maps or albedo by following the same procedures, creating a high quality quantitative regional or global data product.

5. Effects of Image Manipulation Algorithms

[35] Documenting the effects of image manipulation algorithms on quantitative data (i.e., THEMIS radiance data) is an important aspect of algorithm development. In order to assess the validity of the techniques described above, the potential error and corrections they introduce into the final data must be addressed. In this manuscript, we do not attempt to provide an exhaustive error analysis for each of the techniques described previously, but rather we attempt to incorporate reasonable errors and contributions for each technique and view them through the perspective of the mosaic process. Additionally, for nearly all the algorithms discussed in this manuscript a detailed error analysis has been reported in other technical publications that document the algorithms and their effects in detail [e.g., Bandfield et al., 2004; Christensen et al., 2004; Nowicki et al., manuscript in preparation, 2011; Nowicki and Christensen, submitted, 2011].

5.1. Temperature Variation Across the Calibration Flag Correction

[36] As the temperature variation across the calibration flag correction is purely empirical and the subtraction from the image is related solely to image derived parameters that have no corresponding physical measurements, this process should typically only be performed on data where the end product will not be used for spectral analysis. However, this technique further reduces instrument noise, making qualitative spectral variation in decorrelation stretch images significantly more apparent. The temperature differences removed by this technique (e.g., 1–2K, Figure 3) are small enough that it can be safely used if the goal is only derived brightness temperature and not emissivity. The temperature differences are within the predicted temperature error associated with THEMIS images [Christensen et al., 2004]. Subsequently, the atmospheric correction [Bandfield et al., 2004] applies a similar technique (e.g., removing a temperature difference based on image derived parameters); however, this atmospheric correction is rooted in the assumption that THEMIS atmospheric band (∼14.88μm) is not correlated with the surface and the temperature differences calculated from that band are then applied to every other band of the image. This is not the case with the temperature variation across the calibration flag correction algorithm, which is why an extremely large filter (4000 lines) is used to correct the data for each band. If a semi-linear feature is present running parallel to the direction of image acquisition (e.g., a large scarp), it is possible that this technique could induce artifacts (in the form of brightness smearing) into the original data. The likelihood of this is extremely small and occurs infrequently (≪ 0.1% of all images); the potential for artifacts to occur is largely outweighed by the benefits associated with the algorithm.

5.2. Random Noise Removal

[37] The details of the white noise removal algorithm have been documented by Nowicki et al. (manuscript in preparation, 2011) and show that the effect of this algorithm on the order of < ± 1 DN for an individual pixel, though this varies from image to image. Additionally the difference in radiance between any one pixel in the original data and data that have had the white noise removal algorithm applied is typically ≪ 5 × 10−6 W cm−2μm−1 sr−1 (Figure 5) though in the most extreme cases can be as much as 1 × 10−5 W cm−2μm−1 sr−1. In this maximum case, this radiance corresponds to a temperature uncertainty of ∼1.1 K at typical daytime temperatures of 275K. For the former case, the temperature uncertainty is much less, corresponding to ∼0.01K at 275K. The largest of these value differences are always associated with pixels constituting strong thermal boundaries and are easily located and ignored in qualitative and quantitative analyses. From a perspective of a global or regional mosaic product, the final contributions of this algorithm are minimal, as mosaics are typically made from a single band and are displayed as gray scale and not as multiband decorrelation stretch mosaics, where white noise is most evident. Additionally, if overlap between images is present, white noise is further reduced as data are averaged (often several times) in the overlapping regions.

5.3. Image Registration and Image Geometry

5.3.1. Introduction

[38] Image registration is the process that locates images from one data set to other images from the same data set and to additional data sets. Several methods exist for generating well-controlled mosaics and are referred to as controlled (externally referenced mosaics, e.g., through bundle adjustment) and semi-controlled (internally referenced mosaics, e.g., through pixel shifting one image to match another). Completely uncontrolled mosaics result from only a priori information about the camera and spacecraft position, pointing, and camera model.

[39] The absolute accuracy of the camera pointing for THEMIS data is sufficient (less than ±2 pixels or 200m) to allow for the creation of high quality uncontrolled products. There are two main sources of image registration error associated with THEMIS: 1) the uncertainty in the timing of the detector readout as compared to the spacecraft clock count that is recorded for each image, and 2) inaccuracies in the spacecraft pointing/orbital position. Errors related to the timing of the detector readout occur only in the along-track direction and are responsible for approximately ±2 pixels of error. The error is assumed to be Gaussian, meaning that most errors will occur at the sub-pixel to one pixel level, and will thus be un-recognizable in the final mosaic. Errors related to the inaccuracy in the spacecraft pointing/orbital position occur in both the along-track and cross-track directions and are responsible for approximately ±1 pixel error. Since these errors are at the one pixel level, most misregistered images are difficult to identify.

[40] The automatic registration algorithm described above is based off of a single reference image and is characterized as a semi-controlled method; its use can result in systematic offsets, causing the mosaic to be either larger or smaller than predicted. This method is considered semi-controlled because it results in images that are well registered to each other but no attempt has been made to register the data to another data set; in a worst case, a ∼1000km THEMIS mosaic may be >4km (0.4%) larger or smaller than the unregistered mosaic. This size difference results from small inaccuracies in the model that describes the instrument or camera and how images are mapped to the planet. These products are generally well registered to themselves but not necessarily to the surface of the planet. The probability of a single image being misregistered to a large (>1000 images) scene is small (∼1–2%) and in this case nearly all misregistrations are at the 1 to 2 pixel level, though it is often difficult to determine registration at a single pixel level by eye.

[41] Controlled mosaics have the highest level of absolute pixel accuracy on the planet's surface that can be achieved. The procedure for generating a controlled mosaic is difficult and we make no attempt to generate these types of data products using DaVinci; however, ISIS has some methods by which these types of data products can be generated. An established technique and one that ISIS supports is called bundle adjustment [Triggs et al., 2000; Kirk et al., 2006]. In this type of process, images are tied via reference points to both each other and other data sets such as Mars Orbiter Laser Altimeter (MOLA) [Smith et al., 2001]. This process requires the generation of control tie points that represent the same physical place on the planet in both the unreferenced image and the reference data set such as MOLA. Once these locations are established through careful hand examination, the bundle adjustment is performed to generate new camera and spacecraft pointing and positional information. This updated information is used in the map projection step to accurately locate the image to the planet's surface, requiring no further image adjustment or warping after the initial projection.

5.3.2. THEMIS Global Mosaic Registration

[42] For the construction of the THEMIS daytime and nighttime global mosaics, we have elected to forego the use of any image registration algorithm and the mosaics should be considered geographically uncontrolled products. The only information used to project the images on the planet comes from the spacecraft and planet position during image acquisition and the THEMIS camera model. For this study, the data have been dead reckoned, meaning that no offset or warping has been applied to the data other than what is predicted by the initial map projection. By choosing to dead reckon these data, we have foregone the difficult task of tie pointing, relating mosaics to reference locations on the surface, and a bundle adjustment of all images. This task is difficult with THEMIS data for several reasons including: 1) the narrow width and large length of the images, 2) difficulty in automatically identifying suitable tie pointing features, and 3) the number of images required to construct the mosaics (>20,000 individual images for the daytime mosaic, and 18,000 individual images for the nighttime mosaic).

[43] These mosaics are not cartographically controlled but are likely accurate to less than ±2 100m THEMIS pixels on the surface of Mars. This error results from the uncertainty in the timing of the detector readout as compared to the spacecraft clock and inaccuracies in the spacecraft pointing/orbital position. No attempt to correct for these errors has been made and the only information placing the image at the surface location is derived from the spacecraft pointing relative to the planet and the THEMIS camera model. However, poorly registered data or data that reduced the overall quality of the mosaic were removed to produce a high quality product. In practice, very few images (<200 in each global mosaic) were removed due to the above mentioned errors, indicating that these types of errors occur infrequently at levels >1 pixel. Images were removed by visual inspection only when they affected the overall quality of the mosaic, with an end result of high quality mosaics with few and minor registration errors.

5.4. Running Stretches

[44] Classifying the observable effects of running stretches qualitatively is straightforward and functionally results in increased local contrast and detail as compared to standard stretches. Figure 6 illustrates differences associated with a running histogram stretch and a standard histogram stretch. The overall contrast of the image in Figure 6 is no longer controlled by the large bright and dark areas in the scene associated with the valley walls. The difference associated with a running stretch in low contrast areas are often times >15% different from the original stretched values. However, the stretch applied is now not only scene dependent but also dependent on the scale of the features in the scene as related to the box size associated with the running stretch. It is most evident in Figure 6 where ∼500 pixels from the edge of the canyon the contrast of the plains is still relatively low as compared to a scene with just the plains observed. This is a result of the bright and dark areas of the canyon walls dominating the stretch until they are not present in the scene being stretched. This artifact is important to note as it could possibly lead to the misinterpretation of geologic features in a given scene. As running stretches serve to normalize the scene locally, they also prove to be useful in creating large-scale mosaics. If the scene is not locally stretched and long images are included, features just outside of one image, but not another may result in badly mismatched stretches, making mosaicking and blending more difficult.

6. Results

6.1. Large-Scale Seamless Mosaics

[45] One of the main goals of the advanced THEMIS data processing is the ability to create large-scale (i.e., global) mosaics where no seams or individual images are visible. Processing tens of thousands of infrared images acquired at different local times, seasons, and years proves to be a difficult challenge as each of these parameters changes the characteristics of the acquired data. The algorithms described in the previous section, including such techniques as tilt and plaid removal, in addition to the use of running stretches allows the effect of these parameters to be minimized and provide the ability for contrast matching without intensive post-processing and contrast adjustment of the images subsequent to stretching.

[46] In addition to creating large-scale seamless mosaics, this software provides the ability to selectively mosaic images of specified interests. An example of this has been illustrated by Piqueux et al. [2008], where a quantitative temperature mosaic of the south polar cap was constructed to map the composition of exposed ices (e.g., CO2 and H2O). In this case, images were limited by season and surface temperature in order to avoid times when the seasonal CO2 cap was present.

6.2. Mosaicking Processes on Global-Scale THEMIS Data

[47] The THEMIS daytime and nighttime infrared 100m/pixel global mosaics (Figures 7a and 7b) are publicly available through the THEMIS website (, as well as the Java Mission-planning and Analysis for Remote Sensing (JMARS) software package (, and will be released to the Planetary Data System (PDS) as a value added product from the 2001 Mars Odyssey mission.

Figure 7.

The THEMIS global (a) daytime and (b) nighttime mosaics shown in this figure are the combination of the high-frequency information (small-scale) and low-frequency information (large-scale) for viewing at this low-resolution in a simple cylindrical projection. Each box in this figure represents a tile that was individually processed for the construction of the THEMIS daytime and nighttime global mosaic. The equatorial tiles (60°N–60°S) are 30° latitude by 60° longitude with a 2° padding around each tile for blending purposes. The near-polar tiles (60°N–75°N and 60°S–75°S), which were only constructed for daytime data are 15° latitude by 60° longitude with a 2° padding as well. The two polar regions (poleward of 75°), which were also only constructed for the daytime mosaics are 15° latitude by 360° longitude with 2° padding on the anti-poleward edge. These regions were constructed in a polar stereographic projection and have been re-projected to a simple cylindrical projection for this figure.

[48] The mosaicking procedure described in this paper was also employed to create the THEMIS global mosaics, where the same algorithms were utilized following the same steps to process the THEMIS data into a visually appealing, qualitative mosaic. The main differences between the mosaicking of THEMIS global data and the generic procedure is the set of criteria that were used to create the THEMIS daytime and nighttime 100m/pixel global mosaics (section 3.2). Additionally, every image was hand examined to ensure that it was artifact-free and that the location on the planet was relatively accurate (section 5.3.2). This procedure, while eliminating potentially valid data, resulted in a mosaic with significantly higher overall quality than if poorly registered images or low quality data were left in the final product.

[49] While registration errors may cause blurriness and reduce the overall visual appeal of the final product, they are not the only causes of blurriness. Seasonal variations at high latitudes (e.g., poleward of 60°) must also be considered. Several studies [e.g., Bandfield, 2007; Bandfield and Feldman, 2008; Kreslavsky et al., 2008; Smith et al., 2009] indicate that high latitudes likely have large amounts of near surface water ice, which may cause both visual and thermophysical changes in the observed landscape. For instance, Bandfield [2007] illustrates through the use of two THEMIS images obtained at different solar longitudes, it is possible to measure and model temperature differences that are a direct result of sub-surface water ice. Data set observation restrictions, image processing, and mosaicking techniques described in this paper help to minimize this temperature difference; however, temperature variations may occur at different rates throughout the scene depending on the proximity of the ice to the surface [e.g., Bandfield, 2007; Bandfield and Feldman, 2008] resulting in images that have differing physical properties being abutted directly next to one another.

[50] Polar processes may also responsible for large observed changes. Many studies have focused on observed changes associated with the polar regions of Mars, both seasonally [e.g., Leighton and Murray, 1966; Kieffer et al., 2000; Kieffer and Titus, 2001; Bibring et al., 2005] and perennially [e.g., Malin et al., 2001; Byrne and Ingersoll, 2003; Thomas et al., 2005]. In these studies, seasonal changes where the polar cap extends and recedes throughout the year are clearly observed in the temperature data. In this data set, we have minimized these effects through visual inspection to obtain high quality data of the surface and minimize the amount of seasonal CO2 ice present in the mosaic. Changes in the perennial cap have been interpreted through the observations of recession and growth of the various pitted terrain present (e.g., Swiss cheese terrain) [e.g., Malin et al., 2001; Byrne and Ingersoll, 2003]. However, other examples where the polar cap has apparently grown over tens of years are given from differences observed in Mariner 9 to Viking data [James et al., 1979] and in THEMIS data [Piqueux and Christensen, 2008]. These changes are typically small (often <1 pixel) or occur over longer periods of time and are thus insignificant in the final mosaicked product.

[51] Another consideration of these mosaicking techniques is the possibility that non-polar surface features may change. In general, two scales of change are typically observed from orbit, fine scale (meters to 10s of meters) where new deposits, new craters, and modified dunes are typically observed, [e.g., Malin et al., 2006; Bourke et al., 2008] and global-scale changes (10s to 100s of kilometers), [e.g., Christensen, 1988; Smith, 2004; Fenton et al., 2007] which are commonly observed as differences in albedo and thermal inertia. However, changes on the scale of hundreds of meters to kilometers have not yet been observed. The daytime and nighttime THEMIS global mosaics provide excellent means to investigate the possibility of these changes. While this is not the specific goal of these data products, change may be observed as a blurry feature in the mosaics, where there is a difference in the surface observed in two overlapping images that were taken at different times. In many cases, overlapping areas may be separated by several (e.g., >2) Mars years. A key to recognizing this type of change is to distinguish it from poorly registered data. If an image is simply poorly registered, the entire overlapping area should be blurry as a result of the blending algorithm, but if surface features have changed over time, only a small region of the overlapping areas should be blurry. While change effects in mosaics are generally considered artifacts, the techniques described in this paper to normalize and register overlapping images of different times may be used to compare images prior to mosaicking and could be powerful way to search for surface feature changes.

[52] A good application of detecting change would be to search for wind streak changes, where there might be growth or change in direction of the deposit or removal area. At this time we have not yet conclusively identified areas where change has been observed; this remains as a future research opportunity.

6.3. Mosaics From Other Instruments and Additional Data Sets

[53] This software has the ability to ingest and mosaic additional data sets including MOC Wide-angle, CTX, and Viking visible imager data, taking advantage of the advanced data processing techniques originally intended for THEMIS data. Additionally these techniques can be applied to other planetary data sets including the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) [e.g., Yamaguchi et al., 1998].

6.3.1. THEMIS Decorrelation Stretch Mosaics

[54] Here we demonstrate the use of this mosaic and processing software on several different data sets. In order to best demonstrate the techniques described for processing THEMIS data, two mosaics using a running decorrelation stretch radiance mosaic (Figures 8a and 8c) and a single band nighttime temperature mosaic (Figures 8b and 8d) were created. These two mosaics utilize every algorithm described in the methods section. In addition to the fully processed data (Figures 8c and 8d), Figures 8a and 8b are mosaics of the same area and same data, where none of the algorithms described in this paper were applied.

Figure 8.

Before and after images illustrating the effects of algorithms and processing detailed in this paper. (a) THEMIS decorrelation stretch mosaic of bands 8, 7, and 5 (centered at 53.25°E, 21.6°S) illustrating the compositional variation as viewed by THEMIS (15 images, 100m/pixel, 1186 x 1482 pixels, ∼1 min execution time excluding image map projection). This band combination was chosen to highlight the presence of olivine, as there is an olivine absorption in THEMIS band 7. In this case, olivine-bearing materials are concentrated in the floor of the crater. (b) THEMIS unprocessed nighttime temperature mosaic of the same area. (c) THEMIS decorrelation stretch mosaic of the same area. The images in this mosaic have had many of the processes described in this paper applied. (d) THEMIS nighttime temperature mosaic where the images have had many of the processes described in this paper applied. The combination of these two figures illustrates not only the compositional diversity of the area but also the nature of that material. For example, in this case the most olivine enriched materials correlate with the highest thermal inertia values (or warmest nighttime temperatures).

[55] The area chosen is representative of a high thermal inertia crater floor [Edwards et al., 2009], where material with a thermal inertia of >1200 J m−2 K−2 s−2 was identified and has been interpreted as in-place bedrock. Additionally, the decorrelation stretch mosaic of this area demonstrates that this material exhibits the characteristic absorptions associated with olivine-bearing basalts. The use of these two data sets in combination provides the ability to not only establish the composition of the material in question, but also the physical characteristics of the material (e.g., sand sized particles or in situ rock). The use of these two types of data in combination can help reconstruct the geologic history of a region. For example, Rogers et al. [2005, 2009] have utilized this method to characterize the geologic history of several regions on Mars including Ares Vallis [Rogers et al., 2005] and Mare Serpentis [Rogers et al., 2009]. In these instances, compositional as well as thermophysical THEMIS data have been utilized extensively, and indicate that materials with the highest thermal inertia typically correspond to olivine enriched materials. Additional authors have also identified this trend [e.g., Hamilton and Christensen, 2005; Bandfield and Rogers, 2008; Tornabene et al., 2008] observed not only in THEMIS but CRISM and OMEGA data as well.

6.3.2. MOC Wide-Angle Global Data

[56] MOC acquired a complete image of the planet nearly every Mars day (12 Mars Global Surveyor orbits). This data record extends for nearly 5 Mars years. Additionally, Mars Color Imager (MARCI) [Malin et al., 2008] can be mosaicked in a similar manner, providing long-term near continuous coverage of the entire surface of Mars. These data provide a unique view of the surface of Mars and can be used to examine active processes on Mars, such as cloud distributions, (e.g., orographic clouds surrounding volcanoes, low lying clouds filling Hellas basin and Valles Marineris and their respective timings), dust storm activity (e.g., views of the initiation location of dust storms), and additional parameters such as albedo changes related to the re-distribution of dust on the surface. This data complements other global data sets such as Thermal Emission Spectrometer (TES) [Christensen et al., 1998, 2001] global albedo [e.g., Smith et al., 2002; Fenton and Mellon, 2006] and thermal inertia maps [Mellon et al., 2000; Putzig et al., 2005; Putzig and Mellon, 2007]. Figure 9 is an example of a global MOC wide-angle map processed with the mosaic software. An empirical correction for the opposition photometric surge has been applied, though residual effects can still be observed as light toned streaks following the orbit track of Mars Global Surveyor. The photometric opposition surge described above has been observed in Viking data [e.g., Thorpe, 1978, 1979] and has been modeled with physical parameters [e.g., Hapke, 1986; Helfenstein and Shepard, 1999; Hapke, 2002], though these models are complex and the application of these models to imagery often does not completely remove the effect [e.g., Helfenstein et al., 1997; Hapke et al., 1998]. This effect is often most easily corrected by an empirical flat field approach for non-quantitative products.

Figure 9.

By using all the images from a 12 orbit time period (34 images, 6 km/pixel, 3556 × 1779 pixels, ∼30 execution time excluding image map projection) a MOC Wide-angle color image of nearly the entire Martian surface can be created. This mosaic was constructed from orbits 3012 to 3024 of the Mars Global Surveyor spacecraft. These data have been empirically corrected for the opposition phase inversion.

6.3.3. Viking Visible Data

[57] Several mosaics of the summit of Apollinaris Patera utilizing a variety of data sets including Viking (Figure 10), THEMIS daytime and nighttime infrared (Figure 11), and CTX (Figure 12) were created to illustrate the versatility of the mosaicking techniques presented. These mosaics were created in the same projection but at different scales to illustrate the ability of these algorithms to process, register, and mosaic a wide range of data sets.

Figure 10.

An automated mosaic of Viking visual images (7 images, 200 m/pixel, 2665 × 2393 pixels, ∼30 min execution time excluding image map projection) centered on Apollinaris Patera (174°E, 8°S). This example data begins to illustrate the usefulness of combining many different data sets to characterize an area. The white box highlights the location of Figure 11 to provide context.

Figure 11.

Colorized THEMIS nighttime temperature data (301 images, 100 m/pixel, 5336 × 4743 pixels, ∼12 min execution time excluding image map) overlain on a THEMIS daytime temperature mosaic (259 images, 100 m/pixel, 5336 × 4743 pixels, ∼10 min execution time excluding image map projection) centered on Apollinaris Patera (174°E, 8°S). Blue tones correspond to dustier material while red tones correspond to rockier material. This technique of overlaying a colorized nighttime temperature map on a daytime temperature mosaic provides the morphologic context for the thermophysical (nighttime infrared) data. The white box highlights the location of Figure 12 to provide context for the higher-resolution data in Figure 12.

Figure 12.

An automated mosaic of CTX high-resolution (4 images, 6 m/pixel, 20576 × 59705 pixels, ∼2 h execution time excluding image projection) visible images was also created centered on Apollinaris Patera (174°E, 8°S). (a) This is the overview of the mosaic region. Even at this scale, additional details not observed in Figures 10 and 11 can be observed. For example, small radial channels originating at the rim of Apollinaris patera encircle nearly the entire caldera. (b) This illustrates the nature of the layered terrain as discussed in reference to Figure 11. This material typically has an elevated thermal inertia and a fractured appearance. (c) Additionally, other unique features such as the crater with well-defined ejecta is easily observed at the full resolution of CTX data but not in either of the previous data sets. Of additional interest are the small-scale textures associated with the surfaces that have elevated thermal inertia values. These surfaces are typically pitted, have abundant small craters and often have a knobby texture.

[58] The Viking visible imager mosaic was created using seven of the highest resolution data available of the Apollinaris Patera region. These data were projected to 200m/pixel at the latitude of scale (in this case latitude 0°N). The pre-processing steps for these data were performed using the ISIS software package, including image calibration. Once these images are ingested into the mosaic software, contrast matching, blending, and advanced stretching algorithms were applied to create a seamless normalized mosaic. The spacecraft pointing data for these images is quite poor, resulting in image warping and absolute position errors of >100 pixels. This makes auto-registration relatively difficult, though by manually adjusting these data to relatively close locations (e.g., <50 pixels) the auto-registration software was able to find a best match, which corresponded to the best-aligned data. Subsequently, individual images were stretched with a histogram stretch to highlight the regional variations in reflectivity observable in the data, rather than the small-scale variations that would be highlighted with a running histogram stretch.

6.3.4. THEMIS Daytime and Nighttime Infrared Data

[59] Colorizing nighttime temperature data and superposing it on daytime temperature data provides insight into the morphologic features from the daytime data associated with thermophysical properties contributed by the nighttime data. Figure 11 is an illustration of this type of data, where nearly 200 images each were used to create the daytime and nighttime mosaics. As this region is relatively small and close to the landing site of the Mars Exploration Rover Spirit [e.g., Squyres et al., 2004], the coverage of this area with THEMIS data is high and thus many images have been averaged to create these mosaics. In this case (Figure 11), as is the case with most examples of this combination, the material with the warmest nighttime temperatures, and thus highest thermal inertia values, are concentrated where the steepest slopes are observed (e.g., cliffs, walls of craters). However, in this case warmer material is also observed on the flanks of Apollinaris Patera. This may indicate that more rocky material is exposed lower on the slopes of the large volcano. Additionally, relatively rocky material is also visible in the caldera of the volcano. These higher thermal inertia surfaces correspond to fractured and layered regions exposed in the caldera. It is possible that these may be linked to past lava flows which are rockier than the surrounding material.

6.3.5. CTX Visible Data

[60] CTX data of the same area provides high-resolution (∼6 m/pixel) imagery that illustrates the small-scale morphology associated with the Apollinaris Patera caldera. Figure 12 is a mosaic of four CTX images processed through the standard mosaicking techniques described above. Figure 12a shows the overview of the region where the four CTX images were mosaicked together. Figures 12b and 12c are subsequently closer views of this mosaic. In these closer views, more detail regarding the nature of the material is observed. For instance, areas with the highest thermal inertia values (red and yellow tones in Figure 11) correspond to the roughest appearance material in Figure 12. These surfaces have a knobby and bumpy appearance. Small craters are abundant and likely contribute to the elevated thermal inertia values, as the walls of these craters are expected to be composed of less mantled, rockier materials.

[61] The use of these three data sets to provide the overall context, large-scale morphology, thermophysical properties, and small-scale morphology, enables a more complete investigation of the area in question linking large-scale trends with small-scale observations. While the three coincident data sets shown in this example were all qualitative products, the possibility for creating quantitative products exists and may be used to further classify (e.g., linking quantitative compositional THEMIS deconvolution results [e.g., Bandfield, 2008] to quantitative thermal inertia values [e.g., Fergason et al., 2006; Edwards et al., 2008]).

7. Conclusions

[62] We have shown the utility of several advanced image processing techniques as applied to THEMIS and other data sets. However, the possibility of applying these techniques, such as the temperature variation across the THEMIS calibration flag correction, random noise removal, and running stretches, is not limited to the data sets presented in this work and may be applied to past, current, and future data. Though the use of these techniques must be carefully validated for each data set to quantify the effects that may be introduced to the data.

[63] The construction of the mosaics presented and others like them using the techniques described in this paper, provide the ability to view the surface of Mars and geologic problems through many different perspectives. For example, one can obtain high-resolution visible imagery of an area, which may help illuminate compositional and thermophysical data, providing a more complete view of geologic processes on Mars. The mosaicking techniques presented in this work are not limited to the data sets presented here, but may be applied to any data set which can be projected using the USGS ISIS tools.

[64] Additionally, this allows the mosaicking of global planetary data from Mars and other planetary bodies such as new data from the Moon and other outer planets missions. This mosaicking ability provides an unprecedented amount of flexibility to the end user to produce both quantitative and qualitative large-scale seamless products. These products can be constructed relatively easily and scientific investigations are not limited to the extent of a single image; rather, they can be combinations of tens to several thousands of images as is the case with the THEMIS global mosaics.

[65] The data selection, mosaicking procedure, quality control measures, and the registration considerations for the THEMIS daytime and nighttime relative temperature global mosaics are presented here. These mosaics are the highest resolution (100m/pixel) global-scale data sets available for Mars to date.

8. Future Work

[66] Several additional steps may be taken to further improve the THEMIS global mosaic data sets presented here. One major step is to align every THEMIS image on the planet using tie points and a bundle adjustment to Mars Orbiter Laser Altimeter (MOLA) data [Smith et al., 2001]. By using common geographical features (e.g., impact craters) in both THEMIS mosaics and MOLA georeferenced data it is possible to force the THEMIS data to match MOLA through a bundle adjustment, creating a cartographically controlled product that is as geographically accurate as the lower resolution MOLA data allows. Quantitative THEMIS data products could also be created, as the software described does not discriminate between qualitative and quantitative products. The software has the ability to work with numerically meaningful data as well as stretched image values. Possible data products include thermal inertia and 10-band emissivity data mosaics. The difficulty in creating these data products does not lie in the mosaic software but in the initial image processing and calibration. As data processing and calibration techniques improve, more sophisticated THEMIS data products will likely emerge.


[67] The authors would like to thank the JMARS software development team for providing software that aided in the analysis and access to data presented in this work. The authors thank two anonymous reviewers whose careful consideration improved this manuscript. Additionally, the authors would like to thank the THEMIS mission planners and support staff for providing targeting opportunities and helpful discussions and numerous other workers, including Ryan Luk, Michael Veto, Sean Marshall, Taylor Feiereisel, Jordana Friedman, Jessica Kaminski, Lauren Puglisi, and Emily McBryan who aided in the construction of the THEMIS daytime and nighttime global mosaics. This work was funded by the NASA 2001 Mars Odyssey THEMIS project.