2.1.

AVIRIS Data and Analysis

AVIRIS radiance data collected by the Jet Propulsion Laboratory (JPL) on October 14, 2010, at 3-m spatial resolution for the Cuprite, Nevada, site were used as the starting point for the previously reported WV-3 SWIR simulation and mineral mapping.¹² AVIRIS is an imaging spectrometer [hyperspectral imaging (HSI) system], measuring radiance over the 0.4 to $2.5\mu \mathrm{m}$ range in 224 spectral bands at approximately 10-nm spectral resolution and variable spatial resolutions from approximately 2 to 20 m depending on flight altitude.^16,17 Eighty-six SWIR bands between approximately 1.2 and $2.5\mu \mathrm{m}$ with 3-m spatial resolution data were used for this analysis. The AVIRIS data were coregistered to orthorectified Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) imagery,^18,19 and resampled to both 3.7-m native WV-3 resolution, and the proposed WV-3 release spatial resolution of 7.5 m for analysis. The radiance data were corrected to apparent reflectance using the model-based “ACORN” atmospheric correction program²⁰ and further refined using field-measured calibration spectra.²¹ This allows direct comparison to laboratory and field spectra (spectral libraries) using both visual and automated approaches.^22–29 Characteristic mean HSI reflectance signatures for laterally extensive minerals important to understand the general geology and hydrothermal alteration at Cuprite were extracted from the data for known mineral locations [Fig. 2(a)], and mapped using the mixture-tuned-matched filter (MTMF) partial unmixing algorithm^30–32 [Fig. 2(b)].

Fig. 2

(a) Mean shortwave infrared (SWIR) spectra extracted from the airborne visible/infrared imaging spectrometer (AVIRIS) data at 7.5-m spatial resolution for selected image regions of interest (ROI) with known mineralogy. (b) Mixture-tuned-matched filter (MTMF) mineral map showing spectrally predominant mineral for each AVIRIS pixel.¹²

The AVIRIS data were analyzed using a standardized approach—linear transformation utilizing the minimum noise fraction transformation³³ to whiten noise, followed by MTMF.^30–32 Typically used for HSI data, this method locates known spectral signatures in the presence of a mixed or unknown background (Fig. 3). MTMF is also described as “partial unmixing,” as it does not require knowledge of all of the spectral endmembers, thus each individual material can be separately unmixed without knowing the full background composition. MTMF combines the best attributes of the classic matched filter (MF)^34–36 with the feasibility constraints of spectral mixing.³⁷ It allows both determination of specific minerals and estimation of their pixel abundances by calculating two measures: an MF score and infeasibility score (Fig. 3).³²

Fig. 3

MTMF concept showing identification and quantification of a known target spectrum in the presence of a diverse background. Higher MF scores (from 0.0 to $1.0=0\%$ to 100% abundance) indicate that there is more of the target spectrum material in the pixel of interest. Lower infeasibility scores [forming $1\sigma$ and $2\sigma$ (etc.) cones extending from the background to the target] constrain the spectral signature in the context of mixing of the background and target signature. The best spectral matches can be mapped based on meeting the combined criteria of high MF score and low infeasibility score.³²

These two attributes are typically used together to determine occurrence and abundance of a particular material at each pixel of a spectral image dataset. The MF score determines the spectral abundance of the material, while the infeasibility score determines whether the measurement is a feasible mixture of background and the target signature. Two-dimensional scatter plotting is usually used to highlight feasible mixtures; however, this is somewhat subjective. For the purposes of this research, an “MTMF Feasibility Ratio”³⁸ (MF score/infeasibility score) was used for standardization between the datasets. All data analyzed utilizing MTMF were thresholded to show only the best mineral match (one material per pixel) for MTMF feasibility ratios greater than 0.025. AVIRIS results are presented for the spectrally predominant minerals only [Fig. 2(b)] to facilitate later comparison with the simulated WV-3 data. Mineral classes mapped using AVIRIS generally correspond with the main mineral groups and locations shown in previously published alteration maps^5,6 and HSI-derived mineral maps^7–13 (Fig. 1). The same MTMF mapping approach was applied to AVIRIS data and simulated WV-3 data,¹² and subsequently to the on-orbit WV-3 data for consistency.

2.2.

WV-3 Simulation and Results

Simulated WV-3 SWIR bandpass functions were built based on specifications and generic response functions provided by DigitalGlobe (DG) (Table 1).³⁹ AVIRIS reflectance data were spectrally resampled to the WV-3 bands using the simulated bandpasses, which were nominally similar to those measured for the actual WV-3 system launched into orbit in 2014 (Table 1 and Fig. 4).⁴⁰ The spectrally simulated WV-3 data were then resized to 3.7- and 7.5-m spatial resolution using pixel aggregation (averaging). The 7.5-m data were used for subsequent analyses.

Table 1

Comparison of simulated WV-3 shortwave infrared (SWIR) bands versus on-orbit WV-3 SWIR. Full on-orbit WV-3 band responses (measured prelaunch) are shown in Fig. 4.

Fig. 4

WorldView-3 SWIR band responses (September 26, 2013).

Mean SWIR reflectance spectra for selected minerals were extracted from the simulated WV-3 data based on known field locations (same locations as for the AVIRIS data) [Fig. 5(a)]. The spectral signatures were used to perform MTMF to match spectral signatures and map spatial locations [Fig. 5(b)]. Results were visually compared to AVIRIS mean spectra and mapping (Fig. 2).

Fig. 5

(a) Mean reflectance spectra extracted from simulated WV-3 SWIR data at 7.5-m spatial resolution for selected image ROIs with known mineralogy. (b) MTMF mineral map showing spectrally predominant mineral for each simulated WV-3 pixel.¹² Visually compare to Fig. 2 AVIRIS results.

Despite numerous remote sensing campaigns at Cuprite, no fully validated ground truth exists for the site—in fact, HSI data provide the best, most detailed mineral maps available [Fig. 1(b)].¹³ Mineral classes mapped using AVIRIS, as part of this research, generally correspond with the main mineral groups and locations shown in previously published alteration maps^5,6 and HSI-derived mineral maps^7–13 (Fig. 1). For the purposes of this validation, the AVIRIS mineral map described here and shown in Fig. 2(b) was used as surrogate ground truth to evaluate WV-3 performance. These data were directly comparable in spectral and spatial coverage and spatial resolution. Visual comparison of the two mineral maps indicates a good correspondence [Figs. 2(b) and 5(b)]. A confusion matrix⁴¹ computed using the AVIRIS as “ground truth” and evaluated with respect to mapping specific alteration minerals using the simulated WV-3 data,¹² however, shows that the overall accuracy is 50.92% and the per-pixel match between the AVIRIS and WV-3 mineral mapping results using the Kappa coefficient^42,43 is 0.38 (Table 2). The simulated WV-3 data performed best for identification and mapping of buddingtonite, calcite, and silica according to the diagonals of the confusion matrix. Assessment of the errors of commission and omission in Table 2 and the AVIRIS and simulated WV-3 mineral maps illustrate, however, that the simulation mapped a large amount of buddingtonite, where the AVIRIS data did not [Figs. 2(b) and 5(b)]. Table 2 also illustrates that significant portions of the errors for all classes are attributable to unclassified areas, indicating that rather than misidentification of materials, many of the errors are ones where no identification was possible using the multispectral data, presumably because the simulated WV-3 data cannot resolve key spectral features because of lower spectral resolution. Additionally, Table 2 shows that there is significant confusion between alunite, kaolinite, and muscovite as well as between buddingtonite and kaolinite at simulated WV-3 bandpasses and spectral resolution.

Table 2

Confusion matrix and supporting statistics comparing WV-3 simulated results (vertical axis) to airborne visible/infrared imaging spectrometer (AVIRIS) mixture-tuned-matched filter (MTMF) results (horizontal axis). Accuracy 50.92% and Kappa=0.3843.

Excluding unclassified pixels from the confusion matrix analysis improves the overall accuracy to 63.43% and the Kappa coefficient to 0.54 (Table 3). Confusion matrix diagonals are somewhat improved, and errors of commission and omission are more easily evaluated with respect to specific minerals. Buddingtonite, kaolinite, and silica are most often mapped as another mineral. AVIRIS pixels containing alunite, kaolinite, and muscovite are most often missing from the simulated WV-3 MTMF mapping.

Table 3

Confusion matrix and supporting statistics comparing WV-3 simulated results (vertical axis) to AVIRIS MTMF results (horizontal axis). Unclassified omitted. Accuracy 63.43% and Kappa=0.5406.

Some minerals are better separated by the WV-3 SWIR bands than others. Examination of the confusion matrices shows that performance is high for calcite, buddingtonite, and silica. There is still, however, confusion between minerals with similar SWIR spectral features such as alunite, kaolinite, and muscovite, which at high spectral resolution all have characteristic spectral features near $2.2\mu \mathrm{m}$ [Fig. 2(a)].

3.1.

Comparison of SWIR Imagery

SWIR color composites were generated from the AVIRIS data, the simulated WV-3 data, and the on-orbit WV-3 data using bands at approximately 1.65, 2.20, and $2.33\mu \mathrm{m}$ (RGB) (Fig. 6). Potentially altered areas are shown on the images as magenta to red pixels. Based on known mineralogy, the alteration mineral alunite (absorption feature near $2.16\mu \mathrm{m}$) appears dark magenta, whereas areas with argillic (high kaolinite content) alteration and muscovite (absorption features near $2.2\mu \mathrm{m}$) occur as more pure red (pink) colors. Note the close visual match between AVIRIS, simulated WV-3, and on-orbit WV-3 color composites (Fig. 6).

Fig. 6

SWIR color composite images using (a) AVIRIS bands 137, 194, 207 (1.652, 2.198, $2.327\mu \mathrm{m}$) as RGB, (b) WV-3 simulated data using SWIR bands S3, S6, S8 (1.65, 2.20, $2.33\mu \mathrm{m}$) as RGB, and (c) on-orbit WV-3 data using bands S3, S6, S8 (1.661, 2.202, $2.329\mu \mathrm{m}$) as RGB.

3.2.

Comparison of Mean Image Spectra for Known Locations

Mean SWIR reflectance spectra extracted from all three datasets for the previously defined regions of interest (ROI) show a close correspondence between the AVIRIS spectra, the simulated WV-3 spectra, and the actual WV-3 spectra (Fig. 7). All spectra have the expected shapes and show key absorption features, with less than 5% error in all cases. Nearly all of the spectra agree to within about 2%. Clearly, on-orbit WV-3 spectra delineate the mineralogy at Cuprite, exhibiting spectra very similar to those predicted in the WV-3 simulation using AVIRIS HSI data. Specific key features present in the WV-3 spectra include the 2.2 and $2.3\mu \mathrm{m}$ absorption features in muscovite [Fig. 7(a)], the $2.33\mu \mathrm{m}$ feature in calcite [Fig. 7(b)], the asymmetrical $2.2\mu \mathrm{m}$ feature in kaolinite [Fig. 7(c)], the $2.16\mu \mathrm{m}$ feature in alunite [Fig. 7(d)], and the broad hydrothermal silica shoulder feature near $2.25\mu \mathrm{m}$ [Fig. 7(e)]. The buddingtonite feature near $2.1\mu \mathrm{m}$ is still distinct, but is shifted to $2.2\mu \mathrm{m}$ because WV-3 does not include a specific band at $2.1\mu \mathrm{m}$ [Fig. 7(f)]. Together, these spectra form the basis of the MTMF mineral mapping for the Cuprite site.

Fig. 7

Comparison of mean SWIR reflectance spectra extracted from the three datasets (AVIRIS, WV-3 simulated data, and on-orbit WV-3 data) for selected ROIs with known mineralogy: (a) muscovite, (b) calcite, (c) kaolinite, (d) alunite, (e) hydrothermal silica, and (f) buddingtonite. Note that colored symbols indicate the simulated WV-3 spectra, which are nearly identical to the AVIRIS spectra where they overlap, but with fewer bands.

3.3.

Comparison of MTMF Mineral Maps

The on-orbit WV-3 reflectance spectra extracted for the previously defined ROI for selected minerals based on known field locations (Fig. 7) were used to perform MTMF to map the spatial locations of specific minerals [Fig. 8(c)]. This result was visually compared to the AVIRIS MTMF results [Fig. 8(a)] and simulated WV-3 MTMF mapping results [Fig. 8(b)]. The WV-3 on-orbit results [Fig. 8(c)] generally appear to match the simulated WV-3 mapping [Fig. 8(b)]. In some aspects, however, they actually appear to more closely match AVIRIS mapping (better performance than predicted). For example, it appears that the actual WV-3 sensor did a better job of mapping buddingtonite and kaolinite than that shown in the simulated results. Note that buddingtonite occurrences mapped by WV-3 are limited, and mostly correspond to those shown in AVIRIS (cyan colored areas near the center of the left image), while the simulated WV-3 shows more extensive buddingtonite occurrences in areas known to be principally alunite or unclassified (cyan areas near top of center image and along right center edge, Fig. 8).

Fig. 8

(a) SWIR MTMF mineral mapping using AVIRIS, (b) WV-3 simulated data, and (c) on-orbit WV-3 data.

3.4.

Comparative Statistics

Again, the AVIRIS HSI data provide the best, most detailed mineral maps available for the Cuprite site.¹³ Comparison of the AVIRIS and on-orbit WV-3 MTMF mineral maps illustrates that mapped mineral classes generally correspond with main mineral groups and locations shown in previously published alteration maps^5,6 and HSI-derived mineral maps.^7–13 The AVIRIS mineral map shown in Fig. 2(b) and again in Fig. 8(a) provided surrogate ground truth to evaluate the WV-3 performance [Fig. 8(c)]. Confusion matrices were computed using AVIRIS as ground truth and evaluated for mapping each specific alteration mineral. The overall performance of WV-3 on-orbit (Table 4) appears similar to that predicted by the WV-3 simulation (Table 2). While visual comparison of the two mineral maps suggests good correspondence, closer inspection and analysis of the confusion matrices for these data (Table 4) demonstrates an overall accuracy of 55.90% and the per-pixel match between AVIRIS and WV-3 mineral mapping results using the Kappa coefficient^42,43 is 0.39. The on-orbit WV-3 data performed best for identification and mapping of calcite and muscovite according to confusion matrix diagonals. Further assessment of errors of commission and omission indicates that WV-3 mapped a greater amount of buddingtonite compared to AVIRIS. This was less severe for the on-orbit WV-3 mapping, with only 16,263 pixels (2%) mapped as buddingtonite when they should have been unclassified, compared to the simulated WV-3 map of 109,847 pixels (14%) as buddingtonite that were unclassified on the AVIRIS mineral map. On-orbit WV-3 also exhibited considerable errors of commission for kaolinite and alunite mineral mapping. Most errors were caused by mapping unclassified AVIRIS pixels as specific minerals by WV-3 (Table 4). Overall, WV-3 illustrated serious errors of omission with respect to nearly all minerals mapped, indicating that, in many cases, no identification was possible using the multispectral data and the MTMF approach. These errors are similar to those observed in the WV-3 simulation. Table 4 shows that at the actual WV-3 bandpasses and spectral resolution, significant mineral mapping confusion exists between alunite, kaolinite, and muscovite. This is most prevalent in alluvial fan areas, where mixing occurs between all three minerals. Additional confusion exists between buddingtonite and kaolinite, probably due to the fact that WV-3 does not have a spectral band that corresponds exactly to the buddingtonite absorption feature near 2.11 μm. Both results are similar to the previous WV-3 simulation.

Table 4

Confusion matrix and supporting statistics comparing WV-3 results (on-orbit 2014) (vertical axis) to AVIRIS results (horizontal axis). Accuracy 55.90% and Kappa=0.3902.

Exclusion of unclassified pixels from the confusion matrix analysis for on-orbit WV-3 data improves both the overall accuracy to 62.23% and the Kappa coefficient to 0.51 (Table 5). Confusion matrix diagonals are significantly improved and errors of commission and omission are clarified with respect to specific minerals. As observed in the previous WV-3 simulation, buddingtonite, kaolinite, and silica are most often mapped as another mineral. Errors of omission, however, are generally higher for on-orbit WV-3 mineral mapping than indicated by simulation predictions.

Table 5

Confusion matrix and supporting statistics comparing WV-3 results (on-orbit 2014) (vertical axis) to AVIRIS (horizontal axis). Unclassified pixels omitted. Accuracy 62.23% and Kappa=0.5122.

3.5.

Other Approaches and Products

Even though WV-3 is a multispectral sensor, we have compared its performance to the hyperspectral AVIRIS sensor and direct mineral mapping approaches to establish a baseline against HSI mineral mapping. This research establishes that, in fact, WV-3 can directly map some minerals based on their spectral signatures because of the well-designed placement of its SWIR bands. There are also other approaches well-suited to multispectral analysis that can produce unique and useful mapping products. A few selected examples shown here for WV-3 include spectral-based methods such as band ratios, directed color composite images, and mineral indices.

Color ratio composite images have long been used to highlight spectral signature shapes using multispectral data.^4,47–49 These are known to reduce the effects of shade, shadows, and image noise on classification. WV-3 VNIR reflectance ratios, in combination with the WV-3 S3/S5 ($1.66/2.20\mu \mathrm{m}$) ratio, provide legacy capabilities similar to Landsat color ratio composites that include the Thematic Mapper SWIR bands 5/7 ($1.65/2.20\mu \mathrm{m}$) ratio, but at much higher spatial resolution. This ratio generally allows mapping of clays and carbonates.⁴⁹ The well-placed WV-3 SWIR bands also provide a wide range of new possibilities for mapping key materials of interest for geologic and alteration mapping using band ratios. For example, WV-3 SWIR band ratios S5/S6 ($2.16/2.20\mu \mathrm{m}$) and S7/S6 ($2.26/2.20\mu \mathrm{m}$) used together highlight spectral absorption features at $2.2\mu \mathrm{m}$. Band ratio S7/S8 ($2.26/2.33\mu \mathrm{m}$) in combination with band ratio S7/S6 ($2.26/2.20\mu \mathrm{m}$) accentuates the distribution of materials with absorption features near $2.33\mu \mathrm{m}$ and establishes a spectral slope for materials like alunite. Combining S5/S6, S7/S6, and S7/S8 ratios (RGB) as a color composite image allows mapping of materials like muscovite (light red), alunite and related minerals (blue-green), and carbonates (blue) [Fig. 9(a)]. The orange area right–center in the image corresponds to the distribution of hydrothermal silica. Other band ratios can be designed utilizing both VNIR and SWIR bands to highlight different minerals or specific materials, and to identify and/or mask areas of water or vegetation.

Fig. 9

(a) WV-3 SWIR reflectance band ratios S5/S6, S7/S6, S7/S8 ($2.16/2.20\mu \mathrm{m}$, $2.26/2.20\mu \mathrm{m}$, $2.26/2.33\mu \mathrm{m}$) as RGB, (b) selected representative WV-3 mineral spectra, and (c) directed color composite with SWIR bands S5, S6, and S8 (2.16, 2.20, $2.33\mu \mathrm{m}$) as RGB. Color saturation of the right image was enhanced using decorrelation stretching⁵⁰ (hues remain unchanged). The vertical colored lines in the spectral plot (b) show the bands displayed as RGB in the directed color composite image (c). The spectral plots are used to design specific color ratios and directed color composites to show materials of interest in predetermined colors.

Directed color composites are simple three-band color images designed based on the known spectroscopic response of materials of interest. Key materials can be emphasized in specific colors by strategically selecting the three bands used for the RGB display based on spectral features and characteristics of the material(s) of interest. Using bands centered on a specific spectral absorption feature allows design of a color composite showing the desired material(s) in a predetermined color or range of colors. The vertical colored lines in the spectral plot [Fig. 9(b)] show the RGB bands used in the WV-3 directed color composite image of SWIR bands S5, S6, and S8 (2.16, 2.20, $2.33\mu \mathrm{m}$) [Fig. 9(c)]. Image color saturation was enhanced using decorrelation stretching.⁵⁰ Muscovite appears red because of high SWIR band S5 ($2.16\mu \mathrm{m}$) reflectance coupled with low contributions of green and blue caused by absorption in SWIR band S6 ($2.20\mu \mathrm{m}$) and lower reflectance in SWIR band S8 ($2.33\mu \mathrm{m}$). Kaolinite is light magenta because of high contribution from SWIR band S5 ($2.16\mu \mathrm{m}$) in conjunction with some contribution of blue and green from SWIR bands S8 ($2.33\mu \mathrm{m}$) and S6 ($2.20\mu \mathrm{m}$), respectively. Carbonates (calcite) are yellow-green in Fig. 9(c) because of high reflectance in SWIR bands S5 ($2.16\mu \mathrm{m}$) and S6 ($2.20\mu \mathrm{m}$) and low reflectance in SWIR band S8 ($2.33\mu \mathrm{m}$) caused by the carbonate absorption feature near $2.33\mu \mathrm{m}$. Alunite appears blue in the image because of higher reflectance near $2.33\mu \mathrm{m}$ compared to lower reflectance in SWIR bands S5 ($2.16\mu \mathrm{m}$) and S6 ($2.20\mu \mathrm{m}$). Other directed color composites can be designed to highlight specific materials, spectral shapes, or absorption features.

Indices for mapping spectral characteristics of materials using multispectral data have long been used for detection of vegetation (e.g., NDVI).^51–53 They have also been extensively used for geologic/mineral mapping with the ASTER sensor,^48,54,55 which had SWIR spectral bands similar to WV-3, but at a lower (30 m) spatial resolution. Mineral indices are typically more complex images than simple RGB color composites or band ratios. They are designed to show occurrence measures for specific minerals or groups of minerals. There are many possible combinations, several of which are shown in Fig. 10 to illustrate the WV-3 capabilities.

Fig. 10

Selected SWIR mineral indices derived from the on-orbit WV-3 data. (a) WV-3 carbonate index [defined here as WV-3 SWIR bands [$\mathrm{S}6/(\mathrm{S}7+\mathrm{S}8)$], or $2.20\mu \mathrm{m}/(2.26\mu \mathrm{m}+2.33\mu \mathrm{m})$]: areas with carbonate signatures appear red or white. (b) Alunite/kaolinite/pyrophyllite index [defined by the USGS as a relative band depth (RBD) image⁴⁸ using ASTER bands $(\mathrm{S}4+\mathrm{S}6)/\mathrm{S}5$ and implemented here as WV-3 SWIR bands [$(\mathrm{S}3+\mathrm{S}6)/\mathrm{S}5$], or $(1.66\mu \mathrm{m}+2.20\mu \mathrm{m})/2.16\mu \mathrm{m}$]: areas with these minerals are green through white. (c) AL-OH group composition index [defined by CSIRO⁵⁵ as ASTER bands $(\mathrm{S}5+\mathrm{S}7)/\mathrm{S}6$ and adapted here to WV-3 SWIR bands $(\mathrm{S}5+\mathrm{S}7)/\mathrm{S}6$ or $(2.16\mu \mathrm{m}+2.26\mu \mathrm{m})/2.20\mu \mathrm{m}$]: blue is well ordered kaolinite or Al-rich white mica (muscovite, illite, paragonite, pyrophyllite, beidellite). Red is Al-poor (Si-rich) white mica (phengite), or montmorillonite.⁵⁵

These alternative processing and analysis methods demonstrate the wealth of spectral information available in the WV-3 SWIR bands. They illustrate that the WV-3 data are equally amenable to mineral analysis using a wide variety of algorithms and methods developed for multispectral data analysis. In addition to unique SWIR bands, WV-3 provides matchless opportunities for fused VNIR-SWIR analysis and for PAN sharpening of results at an unprecedented 0.31-m spatial resolution.