In the last section of this series, we took a look at both the ASTER sensor and the vast library of spectral profiles built as a standard reference for multi-spectral analysis and material classification. In this article, we will take a look at land cover investigations through spectral unmixing, which is a primary method for classifying materials in remotely sensed data. But to begin with, what exactly is spectral unmixing?
Spectral unmixing may be thought of, in a very general sense, as being similar to what your brain does while listening to a symphony, in that it is able to parse out discreet instruments from the din of horns, strings and percussion all sounding at once. Though an orchestra is trained to play together as perfectly as humanly possible, your ear and brain are still able to differentiate the french horn’s part from the violins; and this is the same sort of process on which spectral unmixing has been developed. The technique rests on the premise that each material has a distinct reflectance spectra to it; and that if we can identify the major materials in a single pixel, we can then unmix their spectra into component parts to get a reasonable sense of what is present. To continue our analogy, it would be like identifying and then removing all of the horns from the symphony so that the strings could be more carefully perceived, without getting “lost” behind the strength of the brass. In the case of spectral unmixing, the instruments are the materials on the ground, the sound is reflected light, and the ear is a multispectral sensor high above the surface of the planet.
There are two models used to describe how reflected spectra (or photons) reach a sensor, and quite typically one is simple and one is not. These models are known as Linear Spectral Mixing and Nonlinear Spectral Mixing. The essence of their difference lies in one basic assumption: Linear Mixing assumes that each photon reflects off only one surface before reaching the sensor; while Nonlinear Mixing allows for a more complex interaction between photons and the materials off which they reflect. A good way to think about it is that Linear Mixing is a macroscopic model (light bouncing once off of a flat roof, for instance) while Nonlinear Mixing accounts for very fine interactions such as is found in soils and sands (e.g. one photon reflecting off of a grain of quartz sand and then a hornblende, and so on). Nonlinear Mixing is relatively new and has not been widely validated, and so, for the sake of concision, we focus here on the more established Linear model.
The first step in Linear Unmixing is the identification of end-members within an image. End-members are those pixels that represent the greatest purity, or the least variance in materials therein. When identifying end-members, it is generally suggested to choose central portions of known bodies such as lakes, parking lots, plowed land, burnt areas and so on, or to identify locations through field work that can be used in end-member classification. Taking ground-truthed spectral readings from field sites can also be incredibly useful for yielding quality results. In a Land Use/Land Cover study of an urban area, for instance, typical end-members would be vegetation, high albedo impervious surface (himp), low albedo impervious surface (limp) and soil (Pu, Gong, Michishita, and Sasagawa, 2008). Himp includes things like concrete, metal roofs and parking lots, while limp includes asphalt, roofing shingles and the like. Soil generally includes bare ground and dry vegetation as well as cultural features without green vegetation. All of these classes may well be identifiable both on the ground and from imagery.
This is when we return to the ASTER spectral library that we covered in the last segment. Not only can the library be used when assigning end-member values, it also becomes valuable in identifying the remaining signals (or materials) that are mixed into the other pixels. When selecting the most pure pixels (or end-members), a semi-supervised method can be used wherein a subsection of the complete library is allocated to perform a sparse solution in the identification of spectrally pure end-members (so termed “sparse” because it makes use of a sparse population of the library). This is an improvement over methodologies that do not make use of a spectral library because it does not rely completely on the availability of pure pixels in an image; and because it reduces reliance on computationally expensive, and often inaccurate, end-member extraction algorithms. The end-members themselves then allow primary unmixing of the remaining pixels with the aid of the spectral profiles in the ASTER library.
Now that we have run through an overview of spectral unmixing and the role of the ASTER spectral library in that process, let’s take a look at an example of both the library and ASTER data being applied to investigate ground cover materials in the Chocolate Mountains of Southern California. Specifically, this study focused on mineral assemblages and the alteration textures associated with below-ground gold deposits. Zhang et al. in 2007 used ASTER data to map the lithology of 1,215 square kilometers of arid desert land, making primary use of the ASTER spectral library to identify the profiles of four alteration minerals from sub-pixel spectral unmixing (i.e. alunite, kaolinite, montmorillonite and muscovite). These particular minerals are important as they tend to form around the same lithologies that are common gold-bearers; and indeed the distributions of these minerals that were identified in this paper were in close association with the units that are most typically gold-bearing. In addition to these alteration minerals, the technique proved capable of mapping a number of geologic units, including volcanic, grantic and metasedimentary formations, with even more specific identification of flood basalts, gneissic units and muscovite schists. The results of this study, which primarily used the 30-meter resolution SWIR and 15-meter VNIR bands on ASTER, are impressive for a few reasons. When compared against existing lithology maps of the study area, the analysis proved to be 84% accurate, with 86% accuracy in gneissic terranes. In addition, the technique correctly identified two bodies of gneiss that had been left off the lithology maps. The corroboration of units that had not been identified on the original maps by field checks of the remotely sensed results is particularly remarkable, and lends a good deal of weight to the inclusion of these spectral techniques in mapping projects of all kinds.
So, how do you, or would you, go about using this technique to learn more about the materials that populate the surface of any planetary body?
About Apollo Mapping
Cameron is a graduate of Pomona College with a degree in Geology, currently working for Apollo Mapping located in Boulder, Colorado. Apollo Mapping is proud to offer Image Hunter, the most fluid search engine of high and medium-resolution satellite data available, with access to all major archive catalog and incredibly fast performance. Apollo is also close to starting a beta test phase of their GISrack, a cloud-based, on-demand, and scalable GIS platform that offers flexible pricing, industry-leading security and performance as well as a large repository of free map data.