# A wavelet-integrated framework for feature extraction and background refinement in hyperspectral anomaly detection

## TL;DR

This paper introduces a new method for detecting anomalies in hyperspectral images using wavelet transforms and other techniques, improving detection accuracy.

## Contribution

The novel contribution is a wavelet-integrated framework (WTHAD) that combines wavelet decomposition, GoDec, and Mahalanobis Distance for improved anomaly detection.

## Key findings

- WTHAD outperforms existing state-of-the-art techniques in hyperspectral anomaly detection.
- Wavelet decomposition helps extract complex signal information and refine background estimation.
- GoDec and Mahalanobis Distance effectively distinguish anomalies from the background.

## Abstract

Wavelet transform (WT) has ability to analyze signals on different frequency bands with resolutions that vary to facilitate the detection of multiple aspects of a signal. With application of Haar WT on each pixel spectrum, the datasets are decomposed into low-frequency approximation coefficients representing essential spectral behaviour and high-frequency detail coefficients. It serves as a powerful preprocessing step for hyperspectral anomaly detection, as anomaly detection is becoming increasingly vital in hyperspectral image analysis. Traditional methods primarily rely on extracting background knowledge and distinguishing anomalies based on their difference from background, which often suffers from challenges such as anomaly contamination. This paper proposes a WT-based method for hyperspectral anomaly detection (WTHAD), which combines WT, Go Decomposition (GoDec) algorithm, and Mahalanobis Distance (MD). It fully exploits WT that offers a powerful preprocessing framework through the analysis of signals in different frequency bands with different resolutions, which facilitates the extraction of complex signal information. GoDec algorithm is used to difference the background and the anomalies. It determines MD in the end to identify the likely anomalies and map anomalies. Six hyperspectral datasets are experimented and it was established that WTHAD has a better performance in detection compared to the existing state-of-the-art hyperspectral anomaly detection techniques.