Jilu Feng

Iterative Spectral Unmixing for Optimizing Per-Pixel Endmember Sets

Fractional abundances predicted for a given pixel using spectral mixture analysis (SMA) are most accurate when only the endmembers that comprise it are used, with larger errors occurring if inappropriate endmembers are included in the unmixing process. This paper presents an iterative implementation of SMA (ISMA) to determine optimal per-pixel endmember sets from the image endmember set using two steps: 1) an iterative unconstrained unmixing, which removes one endmember per iteration based on minimum abundance and 2) analysis of the root-mean-square error as a function of iteration to locate the critical iteration defining the optimal endmember set. The ISMA was tested using simulated data at various signal-to-noise ratios (SNRs), and the results were compared with those of published unmixing methods. The ISMA method correctly selected the optimal endmember set 96% of the time for SNR of 100 : 1. As a result, per-pixel errors in fractional abundances were lower than for unmixing each pixel using the full endmember set. ISMA was also applied to Airborne Visible/Infrared Imaging Spectrometer hyperspectral data of Cuprite, NV. Results show that the ISMA is effective in obtaining abundance fractions that are physically realistic (sum close to one and nonnegative) and is more effective at selecting endmembers that occur within a pixel as opposed to those that are simply used to improve the goodness of fit of the model but not part of the mixture

The topographic normalization of hyperspectral data: implications for the selection of spectral end members and lithologic mapping

Abstract Compact Airborne Spectrographic Imager (CASI) hyperspectral data is used to investigate the effects of topography on the selection of spectral end members, and to assess whether the topographic correction improves the discrimination of rock units for lithologic mapping. A publicly available Digital Elevation Model (DEM), at a scale of 1:50,000, is used to model the radiance variation of the scene as a function of topography, assuming a Lambertian surface. Skylight is estimated and removed from the airborne data using a dark object correction. The CASI data is corrected on a pixel-by-pixel basis to normalize the scene to a uniform solar illumination and viewing geometry. The results show that topography has the effect of expanding end member clusters at times resulting in the overlap of clusters and that the correction process can effectively reduce the variation in detected radiance due to changes in local illumination. When topographic effects are embedded in the hyperspectral data, methods typically used for the selection of end members, such as the convex hull method, can miss end members or result in the selection of nonrepresentative pixels as end members. Thus, end members selected by some conventional methods are very likely “incomplete” or “nonrepresentative” if the topographic effect is embedded in the data. As shown in this study, the topographic correction can reveal hidden end members and achieve a better representation of end members via the statistical center of isolated clusters.

Hyperspectral imaging for the determination of bitumen content in Athabasca oil sands core samples

Ore grade is one of the primary variables controlling the economic recovery of bitumen from oil sands reservoirs, hence there is a need for fast and reliable quantification of total bitumen content (TBC). This is typically achieved through laboratory-based Dean-Stark analyses of drill core samples. However, this method is time and labor intensive and destructive to the core sample. Hyperspectral imaging is a remote sensing technique that can be defined as reflectance spectroscopy with a spatial context, where high-resolution digital imagery (∼1 mm/pixel [0.04 in./pixel]) is acquired and reflectance measurements are collected in each pixel of the image. This study compares two hyperspectral models for the determination of TBC from imagery of both fresh and dry core samples. For three out of four suites of fresh core, TBC was predicted within ±1.5 wt. % of the Dean-Stark data with both spectral models achieving correlations of . For a fourth fresh core and the dry core, larger margins of error were found because of some instances of overestimation. Surface roughness because of uneven oil distribution and small-scale fracturing is a potential source of error in some of the spectral TBC results, particularly for the dry core. Producing results within minutes with the additional benefit of being nondestructive to the core sample, hyperspectral imaging shows great potential to become a viable alternative method for bitumen content determination in oil sands.

Spatial Sub-Sampling Using Local Endmembers for Adapting OSP and SSEE for Large-Scale Hyperspectral Surveys

Airborne and satellite hyperspectral sensors can collect significant quantities of data, commonly terabits in size. Thus, there is a need for the development of computational cost effective algorithms that speed up processing, yet retain essential data quality and information. Endmember extraction is critical in the processing chain of hyperspectral data. Existing methods have primarily been demonstrated on small data sets so their usefulness with large data sets has not been fully explored. The objective of this paper is to adapt the Orthogonal Subspace Projection (OSP) and Spatial Spectral Endmember Extraction (SSEE) algorithms, such that they run efficiently on large data sets without the loss of global image endmember quality. This is demonstrated using a simulated and two real hyperspectral images, the last of which is a multi- flight-line survey. The two adapted methods (OSP SS and SSEE SS) make use of spatial sub-sampling via local endmember extraction to reduce the size of the original data set. This paper will demonstrate that this type of spatial sub-sampling retains the full volume of the data, and thus, the vertices of the simplex that represent the global image endmembers. The results also show that computational time is reduce by more than half using the adapted methods. For the large multi- flight-line survey SSEE SS was able to better retain endmembers of natural materials compared with OSP SS. These results illustrate the potential of spatial sub-sampling for other endmember extraction tools when dealing with multi-flight-line surveys.

Hyperspectral imaging for the characterization of athabasca oil sands drill core

Accurately logging oil sands core is a challenging process because sedimentary and biogenic features are difficult to see in bitumen-saturated strata. A suite of oil sands core was scanned using the SisuROCK hyperspectral imaging system. Analysis of the spectral imagery revealed that in the shortwave infrared we can “see through” thin bitumen at the surface and readily identify structures and fabrics that are not visible to the unaided eye. In the spectral data, we identify broad spectral classes that correlate to lithofacies, which are then used to create rudimentary lithological maps. Greyscale ore-grade maps are also created using an existing spectral model for the determination of bitumen content. Developing an automated method for the removal of man-made materials from the hyperspectral imagery, based on a spectral library developed during this study, remains a research preoccupation. Future work will investigate the discrimination of individual clay minerals, particularly swelling vs. non-swelling.

Rock type classification of drill core using continuous wavelet analysis applied to thermal infrared reflectance spectra

This study investigates a core logging methodology to map rock type using thermal infrared reflectance (TIR) spectra (500–4000 cm–1 or 2.5–20.0 μm) for 74 samples encompassing 11 rock types exposed in various mines of the Sudbury Basin, Canada. A continuous wavelet transform (CWT) was used to represent the original reflectance spectra as a suite of wavelets, each capturing spectral features of different scales with the low-scale components containing mineral spectral features and the high-scale components capturing the overall continuum. Classification was driven by the use of endmember spectra and the spectral angle mapper (SAM). Modelling and validation suites were developed and the mapping accuracy evaluated iteratively for random data splits. The results were compared for reflectance and wavelets of low components of power and significance. We found that the variability amongst measurements observed for varying orientation of a sample or due to variable surface roughness can be greatly minimized with the use of low-scale components, thus improving rock type classification. The average accuracy computed for the 11 rock types is highest for the low-scale component of power (72%) data as opposed to the reflectance data (55%). The highest average accuracy per rock type is obtained using the low-scale components (average value of 82%) for seven rock units that are relatively texturally homogeneous and of uniform modal mineralogy. Lower accuracy values are observed for rock units that display pronounced textural heterogeneity at the scale of observation, or variability in modal mineralogy, or are spectrally similar to other rock types.

Ore detection and grade estimation in the Sudbury mines using thermal infrared reflectance spectroscopy

This pilot study investigated the usefulness of thermal infrared reflectance (TIR) spectroscopy to estimate ore grade in an underground environment and to separate ore‐bearing samples from their host rocks. Work was carried out under laboratory conditions to test the initial concept; all samples had naturally broken faces to mimic the situation in a freshly blasted underground opening. A total of 26 samples, including massive and disseminated ores, were collected from eight mines around the Sudbury basin in Ontario. Rock surfaces were measured wet and dry to address environmental conditions encountered underground. To separate barren rocks from ores and for ore‐grade estimation, an important finding of this research is that, in the region of 1319cm-1, most known silicate minerals converge to a common reflectance minima (<1.5%), but massive and disseminated sulfides have distinctly higher reflectance. Individual sulfide minerals (chalcopyrite, pyrrhotite, pentlandite), however, do not reveal diagnostic fea...

Hyperspectral flight-line leveling and scattering correction for image mosaics

Commonly there is a need to create a “seamless” mosaic for hyperspectral surveys that include multiple adjacent flight-lines. This is done for both visual continuity and to remove line to line radiometric inconsistencies for subsequent analysis. This paper presents a novel empirical approach to generate image mosaics that is two fold. The first step is to apply a per-pixel scattering correction to adjust for BRDF effects. This correction is based on a vegetation index and the mean reflectance of the image. The second step is a leveling procedure that treats each line equally to derive a band by band correction factor that is weighted based on the scan angle of a given pixel. The results show that the scattering correction combined with the leveling procedure can significantly reduces line to line inconsistencies and produce a high quality mosaic.

Predictability of leaf area index using vegetation indices from multiangular CHRIS/PROBA data over eastern China

Abstract. This study analyzed the predictability of leaf area index (LAI) to the variation of vegetation type, observation angle, and vegetation index (VI). The analysis was conducted by using the R2 of the LAI-VI models between in situ measured LAIs and VIs derived from CHRIS/PROBA data. The results show that the discrepancy of vegetation type mostly influences the LAI-VI models. The predictability of LAI to the variation of both vegetation type and index demonstrates the differences of oblique/vertical and backward/forward observations, and backward series are greater than the forward. The predictabilities of LAI to the variation of observation angle are greatest for the soil-adjusted VIs and least for the traditional ratio-based indices. Multivariable linear modeling with all VIs from all five angles yields acceptable accuracy except for the sparse shrub. The backward less-oblique observation (−36  deg) is the only angle chosen in the modeling for grass, shrub, and broad leaf forest, while the nadir view performs best for forests with coniferous trees. These results provide a reference to multiangular LAI estimation for different vegetation communities. VIs accounting for angular soil effects require further investigation in the future.

A spatial–spectral approach for deriving high signal quality eigenvectors for remote sensing image transformations

Spectral decorrelation (transformations) methods have long been used in remote sensing. Transformationof the image data onto eigenvectors that comprise physically meaningful spectral properties (signal) canbe used to reduce the dimensionality of hyperspectral images as the number of spectrally distinct signalsources composing a given hyperspectral scene is generally much less than the number of spectral bands.Determining eigenvectors dominated by signal variance as opposed to noise is a difficult task. Problemsalso arise in using these transformations on large images, multiple flight-line surveys, or temporal datasets as computational burden becomes significant. In this paper we present a spatial–spectral approachto deriving high signal quality eigenvectors for image transformations which possess an inherently abil-ity to reduce the effects of noise. The approach applies a spatial and spectral subsampling to the data,which is accomplished by deriving a limited set of eigenvectors for spatially contiguous subsets. Thesesubset eigenvectors are compiled together to form a new noise reduced data set, which is subsequentlyused to derive a set of global orthogonal eigenvectors. Data from two hyperspectral surveys are used todemonstrate that the approach can significantly speed up eigenvector derivation, successfully be appliedto multiple flight-line surveys or multi-temporal data sets, derive a representative eigenvector set forthe full image data set, and lastly, improve the separation of those eigenvectors representing signal asopposed to noise.