Improved manifold coordinate representations of hyperspectral imagery

Abstract

Abstract-There are many well-known sources of nonlinearity present in hyperspectral imagery; these include bi-directional reflectance distribution function (BRDF) effects, multi-path scatter between heterogeneous pixel constituents, and the variable presence of water, an attenuating medium, in the scene. In recent publications, we have presented a data-driven approach to representing the nonlinear structure of hyperspectral imagery [4]. The approach relies on graph methods to derive geodesic distances on the high-dimensional hyperspectral data manifold. From these distances, a set of manifold coordinates that parameterizes the data manifold is derived. Because of the computational and memory overhead required in the geodesic coordinate calculations, the approach relies on partitioning the scene into subsets where the optimal manifold coordinates can be derived in an efficient manner, followed by an alignment stage during which the embedded manifold coordinates for each subset are aligned to a common manifold coordinate system. In [4], we demonstrated the feasibility of the coordinate and alignment methodology and the ability of the manifold approach to provide higher data compression and more effective classification when compared with linear methods. In this paper we develop an improved approach to the manifold coordinate alignment phase with an improved sampling methodology. Results are demonstrated using examples of hyperspectral imagery derived from PROBE2 hyperspectral scenes of the Virginia Coast Reserve barrier islands.

Efficient Boundary Tracking Through Sampling

“…The Smith Island H20000508 southendBIP data set [8,5,6,7] is a large hyperspectral data set obtained by an airborne hyperspectral scanner (HyMAP) on May 8, 2000, over Smith Island, VA, a barrier island in the Virgina Coast Reserve. The boundary between land and sea is often ambiguous, causing problems especially when the full spectral information is not used.…”

“…7. The Smith Island data set [8,5,6,7]. Left: Reference points used in the land/water classification by full unmixing.…”

Manifold coordinate representations of hyperspectral imagery: Improvements in algorithm performance and computational efficiency

“…Working with the Naval Research Laboratory through a Cooperative Research and Development Agreement (CRADA), Celestech has implemented the NRL manifold representation algorithm described in [2]. On the GPU, the computation speed of this algorithm has increased by 2 orders of magnitude over optimal CPU-based implementations.…”

“…Figure 6. GPU acceleration of the three processing stages used to obtain manifold coordinates using the algorithm described in [2]. The reconstruction step (stage 3) shows the most dramatic acceleration compared to processing on a conventional workstation.…”

“…This scalable approach to deriving an intrinsic coordinate system for the data has allowed us to develop a practical framework for processing large-scale remote sensing imagery in a variety of applications such as water column retrievals, e.g. bathymetry and bottom type [3], as well as land-cover [2] and anomaly finding [4] [8].…”