E2E-LIADE: End-to-End Local Invariant Autoencoding Density Estimation Model for Anomaly Target Detection in...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral anomaly detection (HAD) identifies atypical spectral samples.
• Existing density estimation methods face challenges in representation learning and preserving spectral information in lower dimensions.
• Two-stage optimization and inability to learn discriminative low-dimensional representations hinder performance.

Purpose of the Study:

• To propose a novel end-to-end local invariant autoencoding density estimation (E2E-LIADE) model for hyperspectral anomaly detection.
• To address limitations of existing methods by integrating representation learning and density estimation.
• To improve the accuracy and efficiency of identifying anomalous targets in hyperspectral data.

Main Methods:

• Introduced a local invariant autoencoder (LIA) to capture intrinsic low-dimensional manifolds.
• Generated augmented low-dimensional representations (ALDR) by combining local invariant features and reconstruction error.
• Employed an end-to-end (E2E) multidistance measure (MSE, OPD) and simultaneously optimized ALDR and density estimation network.

Main Results:

• The E2E-LIADE model achieved superior performance compared to state-of-the-art methods.
• Simultaneous optimization prevented local optima and generated an effective energy map.
• Postprocessing refined the energy map for enhanced background suppression and target detection.

Conclusions:

• The proposed E2E-LIADE model offers a significant advancement in hyperspectral anomaly detection.
• The end-to-end approach effectively learns discriminative representations and optimizes density estimation.
• The method demonstrates robust performance in identifying anomalous targets within hyperspectral imagery.