Milton O. Smith

Classification of multispectral images based on fractions of endmembers: Application to land-cover change in the Brazilian Amazon

Abstract Four time-sequential Landsat Thematic Mapper (TM) images of an area of Amazon forest, pasture, and second growth near Manaus, Brazil were classified according to dominant ground cover, using a new technique based on fractions of spectral endmembers. A simple four-endmember model consisting of reflectance spectra of green vegetation, nonphotosynthetic vegetation, soil, and shade was applied to all four images. Fractions of endmembers were used to define seven categories, each of which consisted of one or more classes of ground cover, where class names were based on field observations. Endmember fractions varied over time for many pixels, reflecting processes operating on the ground such as felling of forest, or regrowth of vegetation in previously cleared areas. Changes in classes over time were used to establish superclasses which grouped pixels having common histories. Sources of classification error were evaluated, including system noise, endmember variability, and low spectral contrast. Field work during each of the four years showed consistently high accuracy in per-image classification. Classification accuracy in any one year was improved by considering the multiyear context. Although the method was tested in the Amazon basin, the results suggest that endmember classification may be generally useful for comparing multispectral images in space and time.

Vegetation in deserts: I. A regional measure of abundance from multispectral images

Abstract A method was tested in the semiarid Owens Valley, California for measuring sparse vegetation cover using Landsat Thematic Mapper (TM) multispectral images. Although green vegetation has a characteristic reflectance spectrum in the visible and near-infrared, using conventional image-processing methods, it has been difficult to quantify vegetation cover of less than about 40%, owing to the spectral dominance of the background soils and rocks. Thus multispectral images have been of limited use in mapping variations in vegetation cover in arid and semiarid regions. In this study fractions of vegetation, soils, and shading and shadow within the smallest resolution elements (30 × 30 m pixels) of the TM images were computed by applying a mixing model based on laboratory and field reference spectra. Fractions of vegetation were calculated for each pixel in TM images taken in December 1982 and May 1985, and the results were compared with ground transects. Despite spatial variations in background soil, temporal differences in satellite instrument response, and differences in atmospheric and lighting conditions, the fractions of vegetation computed from each image gave a spatially consistent measure of the projected vegetation cover. Results were obtained for a 150-km segment of Owens Valley; they indicate that the method can facilitate mapping and monitoring sparse vegetation cover over large regions covered by satellite images.

Green vegetation, nonphotosynthetic vegetation, and soils in AVIRIS data

Abstract An Airborne Visible / Infrared Imaging Spectrometer (AVIRIS) image collected over the Jasper Ridge Biological Preserve, California on 20 September 1989 was analyzed using spectral mixture analysis. The scene was calibrated to reflectance assuming a homogeneous atmosphere. The image was modeled initially as linear mixtures of the minimum number of reference endmember spectra that accounted for the maximum spectral variability. Over 98% of the spectral variation was explained by linear mixtures of three endmembers: green vegetation, shade, and soil. Additional spectral variation appeared as residuals. Nonlinear mixing was expressed as variations in the fraction of each endmember when a linear mixing model was applied to spectral subsets of the entire spectrum. After the fractions of the endmember spectra were calculated for each pixel, different types of soil were discriminated by the residual spectra. Nonphotosynthetic vegetation (NPV) (e.g., dry grass, leaf litter, and woody material), which could not be distinguished from soil when included as an endmember, was discriminated by residual spectra that contained cellulose and lignin absorptions. Distinct communities of green vegetation were distinguished by 1) nonlinear mixing effects caused by transmission and scattering by green leaves, 2) variations in a derived canopy-shade spectrum, and 3) the fraction of NPV. The results of the image analysis, supported by field observations in 1990 and 1991, indicate that the multiple bands of AVIRIS enhance discrimination of NPV from soil, and the separation of different types of green vegetation. The ability of the system to measure narrow absorption bands is one important factor; however, also important is the variation in continuum spectra expressed by the endmembers, and characteristic nonlinear mixing effects associated with green leaves.

Quantitative subpixel spectral detection of targets in multispectral images

Spectral mixture analysis was used to determine threshold detection limits of target materials in the presence of background materials within the field of view under various simulated but realistic compositional, instrumental, and topographical conditions. Detection thresholds were determined for the cases where the target is detected as (1) a component of a spectral mixture (continuum threshold analysis) and (2) residuals (residual threshold analysis). In continuum threshold analysis, the target was included as a component during unmixing thereby permitting evaluation of target detectability. In residual threshold analysis, the unmodeled target was detected as wavelength-dependent deviations of the spectral mixture (target included) from the predicted spectrum (mixtures of the modeled background spectra). High resolution laboratory spectra were used to test the “best case” for target detection in spectral mixtures. Data quality was then decreased to simulate the effects of various imaging instruments (spectral sampling and noise) and changes in lighting geometry. The results show that the contrast of the target spectrum, relative to mixtures of background spectra, determines which level of spectral mixture analysis (continuum or residual analysis) detects the target at the lower threshold. Continuum analysis provides lower thresholds when there is contrast between the target and mixtures of the background spectra throughout much of the spectrum, whereas residual analysis improves detectability when the uniqueness of the target is found in narrow absorption feature(s). Varying spectral sampling (by wavelength) changes the contrast (and detectability) of the target and background materials. Although simple examples of target and background materials are used in this paper, they illustrate a general approach for evaluating the spectral detectability of terrestrial and planetary targets at the subpixel scale. Using spectral mixing analysis as a framework makes it possible to specify the conditions that are necessary to detect a particular material and the most appropriate imaging system for a given application.

Estimating suspended sediment concentrations in surface waters of the Amazon River wetlands from Landsat images

A method has been developed, based on spectral mixture analysis, to estimate the concentration of suspended sediment in surface waters of the Amazon River wetlands from Landsat MSS and TM images. Endmembers were derived from laboratory reflectance measurements of water-sediment mixtures with a range of sediment concentrations. Using these reference spectra, we applied a linear mixture analysis to multispectral images after accounting for instrument and atmosphere gains and offsets. Sediment concentrations were estimated for individual pixels from the mixture analysis results based on a nonlinear calibration curve relating laboratory sediment concentrations and reflectance to endmember fractions. The uncertainty in the sediment concentrations derived from this analysis for three Amazon images is predicted to be within ±20 mg/L, and the concentrations fall within a range of concentrations of suspended sediment that were measured at several times and places in the field over the past 15 years. The emphasis of our work is to use the patterns of sediment concentrations to compute the approximate volumes of sediment that are transferred between the main channel and floodplain of the Amazon River. However, the methodology can be applied universally if the optical properties of water and sediment at the site are known, and it is, therefore, useful for the study of suspended sediment concentrations in surface waters of wetlands elsewhere.

Simple Models For Complex Natural Surfaces: A Strategy For The Hyperspectral Era Of Remote Sensing

A two-step strategy for analyzing multispectral images is described. In the first step, the analyst decomposes the signal from each pixel (as expressed by the radiance or reflectance values in each channel) into components that are contributed by spectrally distinct materials on the ground, and those that are due to atmospheric effects, instrumental effects, and other factors, such as illumination. In the second step, the isolated signals from the materials on the ground are selectively edited, and recombined to form various unit maps that are interpretable within the framework of field units. The approach has been tested on multispectral images of a variety of natural land surfaces ranging from hyperarid deserts to tropical rain forests. Data were analyzed from Landsat MSS (multispectral scanner) and TM (Thematic Mapper), the airborne NS001 TM simulator, Viking Lander and Orbiter, AIS, and AVRIS (Airborne Visible and Infrared Imaging Spectrometer).

Structural stage in Pacific Northwest forests estimated using simple mixing models of multispectral images

Abstract We identified stages of regrowth in replanted clearcuts in Douglas-fir/western hemlock forests in the Gifford Pinchot National Forest, southern Washington, USA, using a simple four-endmember constrained linear spectral mixing model applied to a multispectral Landsat Thematic Mapper image in order to separate and quantify spectral contributions from significant scene components. Spectral unmixing produces images of the fractional amount of the spectral endmembers, which were green vegetation, nonphotosynthetic vegetation, soil and “shade,” which includes topographic shading and shadows. Changes in endmember fractions correspond to changes in surface composition (as viewed from above). Unresolved shadows comprise the primary indicator of canopy structure and hence, regrowth stage. To isolate shadows, shading predicted from a digital elevation model was removed from the image before mixture analysis. As stands regrow, the surface cover shifts from initial high proportion of slash and exposed soil, and low proportions of green vegetation and shadows, to low fractions of stems and soil with high fractions of green vegetation and shadows. This shift in surface composition defines a regrowth trend in an endmember fraction data space. Projection of data onto this line allows estimation of structural stage and stand age, and provides a framework for remote mapping and monitoring of forest regrowth. Field analysis of 495 forest stands, representing stand structural stages ranging from newly replanted cuts to stands greater than 250 years in age, was used to assess the accuracy and precision of predicted structural stages and stand ages. The spectral unmixing approach can be used to evaluate and monitor forest regrowth quickly over large areas of the Pacific Northwest forests, and is extendible to mapping basic vegetation community type as well as structural stage.

Simple algorithms for remote determination of mineral abundances and particle sizes from reflectance spectra

Simple algorithms for quantitatively modeling the reflectance spectra of mineral particulates are tested. Although more sophisticated models exist, these algorithms are particularly suited for remotely sensed data, where little or no opportunity exists to independently measure reflectance versus particle size and phase function. Previously, we introduced this method in the analysis of the directional-hemispherical reflectance spectra of binary mineral mixtures containing a single particle size distribution. In this study the technique is extended to multicomponent mixtures, various size separates, and spectra with differing illumination/viewing geometries. It is found that the theoretical calculations and measured data agree to nearly the level of experimental error. This method is also used to determine the threshold abundance at which a mineral can be detected when mixed with another mineral.

Vegetation in deserts: II. Environmental influences on regional abundance

Abstract A remote-sensing approach was used in conjunction with field measurements to examine local and regional-scale environmental processes that covary with the abundance and distribution of vegetation in a semiarid ecosystem. Images of the fractional abundances of vegetation and soils were constructed by spectral mixture analysis of Landsat Thematic Mapper (TM) satellite images, covering a 150-km segment of Owens Valley, California. These images, along with a TM image of the radiant temperature, a digital elevation image and ground-based measurements of precipitation and evapotranspiration, were examined to isolate the effects on vegetation of the covarying factors, net radiation, temperature, elevation, soil type, and precipitation. On a regional scale the abundance of desert scrub on the bajadas of Owens Valley appears to be influenced most by the mean annual precipitation. Also regionally, vegetation cover is sensitive to the differences between the gravelly fanglomerates of the bajadas and the alluvium of the valley floor. Other edaphic and ground-water effects are important but localized, and are most pronounced on the valley floor. They produce patterns in vegetation abundance that are independent of and superposed on the regional precipitation-controlled pattern. Temperature covaries with vegetation less well than precipitation, and appears not to be the major influence on either the amount of vegetation or the boundaries between major vegetation communities. The image-derived measure of vegetation cover correlates closely with ground-based measurements of evapotranspiration. The study demonstrates that local observations cannot be extrapolated safely to the regional scale, and that a combination of local field measurements and the regional measurements provided by remote sensing is required to determine the environmental factors that control vegetation distribution.

SPECTRAL MIXTURE ANALYSIS - NEW STRATEGIES FOR THE ANALYSIS OF MULTISPECTRAL DATA

Instrument noise, spectral contrast among scene components and variability of spectral scene components are not explicitly evaluated as part of classification and mapping efforts using multispectral images. Yet changes in these factors directly affect mapping accuracy. An analytical framework is proposed such that these factors can be quantified within the context of spectral mixture analysis (SMA). In applying these analyses to an AVBRIS image of Owens Valley, California, U.S.A., we find that the greatest uncertainty in abundance estimates arises from spectral variability in endmembers. Spectral variability in any endmember results in abundance uncertainty of all endmembers. We propose an analytical strategy that subsets an image into regions of lowest spectral dimensionality to minimize uncertainties and to maximize detection of new materials.