Automatic classification of land cover on Smith Island, VA, using HyMAP imagery

Abstract

Abstract-Automatic land cover classification maps were developed from Airborne Hyperspectral Scanner (HyMAP) imagery acquired May 8, 2000 over Smith Island, VA, a barrier island in the Virginia Coast Reserve. Both unsupervised and supervised classification approaches were used to create these products to evaluate relative merits and to develop models that would be useful to natural resource managers at higher spatial resolution than has been available previously. Ground surveys made by us in late October and early December 2000 and again in May, August, and October 2001 and May 2002 provided ground truth data for 20 land cover types. Locations of pure land cover types recorded with global positioning system (GPS) data from these surveys were used to extract spectral end-members for training and testing supervised land cover classification models. Unsupervised exploratory models were also developed using spatial-spectral windows and projection pursuit (PP), a class of algorithms suitable for extracting multimodal views of the data. PP projections were clustered by ISODATA to produce an unsupervised classification. Supervised models, which relied on the GPS data, used only spectral inputs because for some categories in particular areas, labeled data consisted of isolated single-pixel waypoints. Both approaches to the classification problem produced consistent results for some categories such as Spartina alterniflora, although there were differences for other categories. Initial models for supervised classification based on 112 HyMAP spectra, labeled in ground surveys, obtained reasonably consistent results for many of the dominant categories, with a few exceptions. For an invasive plant species, Phragmites australis, a particular concern of natural resource managers, this approach initially had an excessively high false-alarm rate. Increasing the number of spectral training samples by an order of magnitude and making concomitant improvements to the georectification led to dramatic improvements in this and other categories. The unsupervised spatial-spectral approach also found a cluster closely associated with Phragmites patches near the thicket boundary, but this approach did not identify the exposed are compared against HyMAP image spectra at model-predicted pixels and at validated GPS waypoints.Index Terms-Barrier islands, hyperspectral, in situ spectrometry, invasive plant species, land cover classification, neural networks, principle component analysis, projection pursuit, supervised classification, unsupervised classification.

Landscape position and habitat polygons in a dynamic coastal environment

“…Ground based reflectance measurements were used to calibrate the imagery data based on target end members collected with an ASD FieldSpec Pro Full range reflectance radiometer (Analytical Spectral Devices Inc., Boulder, Colorado, USA). Classification of individual plant species followed methods demonstrated with coastal vegetation (Bachmann et al 2002, Bachmann 2003. Field reflectance spectra for individual species were collected and wavelength bands identified that distinguished species.…”

Improved manifold coordinate representations of hyperspectral imagery

“…Sources of nonlinearity include: (1) nonlinear variations in reflectance produced by variations in sun-canopy-sensor geometry in the landscape [15] [9], (2) multi-path scatter among sub-pixel constituents [12] [14], violating the traditional linear mixing assumptions, (3) the variable presence of water, an attenuating medium [11] in the scene. Some of the errors that we observed in mapping products that we previously derived in [2] [3] became the motivation for finding new methods of modeling nonlinear structure in hyperspectral data [4]. In the next two subsections, we give a brief overview of the approach that we presented in [4] as a preamble to introducing improvements.…”

“…Note that the definition of "tractable" was addressed in [4], but improved scaling presented in Section II expands the scale of what is considered a computationally feasible tile size. The ISOMAP portion of the computations involves the following steps: (1) given a specified metric such as Euclidean, spectral angle, or some other appropriate choice, determine the spectral neighborhoods (initial sparse neighborhood graph) where linearity holds, maintaining a list for each sample of its neighbors and metric distances; (2) at each sample, for all distances outside the neighborhood, use Dijkstra's algorithm [7] [16] with a minimum priority queue to relax the closest edge not already attached to the graph d G to compute the shortest nonlinear path (geodesic) distance to all other samples (note that this is a graph calculation and that the metric, therefore, is not involved here but is only evaluated in step (1) inside the neighborhood); (3) if there are any remaining distances which can not be connected in the distance graph, attach pockets of isolated points to each other in the graph by find the closest linear distance between pairs of isolated pockets, thus preserving the geodesic structure of each and ensuring a minimal spanning tree [4] 1 ; after attaching all points symmetrize to ensure consistency of paths in both directions; (4) with the full NxN (N=number of spectral samples) geodesic distance matrix calculated in steps (1) and (2), compute the second order variation in the geodesic distances,…”

“…[8] began acquiring imagery over the island chain twice annually in spring and summer. One focus of our research at the Virginia Coast Reserve site has been the devlopment and testing of algorithms for detailed species-level mapping of coastal land-cover [2] [3] [1] [4]. Using the hyperspectral time series, we have developed fast online methods for fusing the clasification results from multi-temporal inputs, for example, mapping products developed for different seasons [3].…”

“…The fusion of classifier results uses smooth estimated measures of classifier reliability to determine a final category at each pixel. Extensive ground truth data has been collected by us [2] [3] throughout all of the islands in the chain, including both in situ reflectance data, BRDF, and more recently biophysical data including biomass, canopy light penetration, leaf area index (LAI), and leaf optical measurements [5].…”