JPSA

JPSA performs linearized subspace analysis for hyperspectral dimensionality reduction (HDR) by jointly learning spatial–spectral manifold-aligned latent subspaces and a linear classifier within a semisupervised framework.

Key Features:

• Linearized Subspace Analysis: Uses a linear approach to mitigate limitations of nonlinear subspace techniques such as poor explainability, high computational cost, limited generalization capability, and inadequate spatial–spectral discrimination.
• Spatial–Spectral Manifold Alignment: Aligns manifold structures within each latent subspace to preserve topological properties between the original and compressed data.
• Joint Learning of Latent Subspaces and Linear Classifier: Simultaneously learns latent subspaces and a linear classifier to identify discriminative projection directions for improved classification.
• Progressive Search for Optimal Mapping: Progressively searches through intermediate subspace states to refine the mapping from the original data space to a more discriminative subspace.
• Semisupervised Learning Framework: Leverages both labeled and unlabeled data to enhance robustness and generalization when labeled samples are limited.

Scientific Applications:

• Hyperspectral Dimensionality Reduction: Reduces high-dimensional hyperspectral data while preserving spatial–spectral information and topological properties.
• Land Cover Classification: Produces discriminative features for land cover mapping from hyperspectral imagery.
• Environmental Monitoring: Supports analysis tasks that require preserved spatial–spectral signatures over time and space.
• Resource Exploration: Facilitates identification of spectral signatures relevant to mineral and resource detection.

Methodology:

Applies linearized subspace analysis with spatial–spectral manifold alignment, jointly learns latent subspaces and a linear classifier via a progressive search through intermediate subspace states, operates in a semisupervised framework using labeled and unlabeled samples, and was validated on the Indian Pines and University of Houston hyperspectral datasets using a nearest neighbor (NN) classifier.