Nonlinear Dynamic Field Embedding: On Hyperspectral Scene Visualization

Graph embedding techniques are useful to characterize spectral signature relations for hyperspectral images. However, such images consists of disjoint classes due to spatial details that are often ignored by existing graph computing tools. Robust parameter estimation is a challenge for kernel functions that compute such graphs. Finding a corresponding high quality coordinate system to map signature relations remains an open research question. We answer positively on these challenges by first proposing a kernel function of spatial and spectral information in computing neighborhood graphs. Secondly, the study exploits the force field interpretation from mechanics and devise a unifying nonlinear graph embedding framework. The generalized framework leads to novel unsupervised multidimensional artificial field embedding techniques that rely on the simple additive assumption of pair-dependent attraction and repulsion functions. The formulations capture long range and short range distance related effects often associated with living organisms and help to establish algorithmic properties that mimic mutual behavior for the purpose of dimensionality reduction. The main benefits from the proposed models includes the ability to preserve the local topology of data and produce quality visualizations i.e. maintaining disjoint meaningful neighborhoods. As part of evaluation, visualization, gradient field trajectories, and semisupervised classification experiments are conducted for image scenes acquired by multiple sensors at various spatial resolutions over different types of objects. The results demonstrate the superiority of the proposed embedding framework over various widely used methods.

💡 Research Summary

The paper tackles the long‑standing problem of visualizing and reducing the dimensionality of hyperspectral images (HSI), which contain hundreds of spectral bands per pixel and exhibit complex spatial‑spectral relationships. Traditional graph‑based embedding methods such as t‑SNE, ISOMAP, LLE, and Laplacian Eigenmaps typically rely on a single similarity measure (often purely spectral) and ignore the spatial context that often separates disjoint classes. Consequently, these methods struggle to preserve meaningful neighborhoods when classes are interleaved or when intra‑class spectral variability is high.

Key Contributions

1. Spatial‑Spectral Composite Kernel – The authors propose a new kernel that jointly incorporates normalized spectral vectors and 2‑D spatial coordinates. By weighting or multiplying exponential terms for spectral distance and spatial distance, the kernel can be tuned (via automatically estimated parameters) to emphasize either modality as needed. This results in a graph where edges are strong only when pixels are both spectrally similar and spatially proximate, thereby preventing spurious connections across distant but spectrally similar regions.

2. Force‑Field‑Based Nonlinear Embedding Framework – Inspired by mechanics, each data point is treated as a particle subject to an attraction force (short‑range, preserving local topology) and a repulsion force (long‑range, encouraging global dispersion). The attraction function is typically a Gaussian‑like decay, while the repulsion follows an inverse‑power law. The total force on particle i is the additive sum over all other particles:
   \