Band selection-based collaborative representation for anomaly detection in hyperspectral images

Donglong Zhu, Bo Du, Liangpei ZhangNational 2020 · DOI

Abstract

Hyperspectral imaging (HSI) provides high spectral resolution, enabling the detection of subtle differences between materials. However, the high dimensionality and inherent redundancy of HSI data often hinder the effective separation of anomalous targets from complex backgrounds. This paper introduces a novel Band Selection-based Collaborative Representation (BSCR) algorithm designed to mitigate these issues. By employing an Optimal Clustering Framework (OCF) for intelligent band selection, the method enhances the separability of anomalies. Subsequently, a Collaborative Representation (CR) model is used to reconstruct pixels, where anomalies yield higher residual errors, facilitating their detection. Experimental results across three diverse datasets demonstrate that BSCR achieves superior Area Under Curve (AUC) values compared to traditional and contemporary state-of-the-art detectors.

1. Introduction

Hyperspectral sensors capture data across hundreds of narrow, contiguous spectral bands, providing a unique "fingerprint" for each pixel. This capability is particularly advantageous for anomaly detection, where the goal is to identify rare objects that deviate significantly from their local surroundings. Despite this advantage, the "curse of dimensionality" presents a significant challenge: many bands are highly correlated or contain noise, which can obscure the subtle signals of interest.

Traditional approaches often rely on dimensionality reduction techniques like Principal Component Analysis (PCA). While effective at reducing data volume, PCA can inadvertently smooth out the very anomalies it seeks to find by transforming the data into a new coordinate space. To address this, the proposed BSCR algorithm prioritizes the preservation of original spectral signatures through intelligent band selection.

2. Methodology

The BSCR algorithm operates in two primary stages: Optimal Band Selection and Collaborative Representation Detection.

2.1 Optimal Band Selection

Instead of transforming the spectral space, BSCR selects a subset of the most informative original bands. This is achieved through the Optimal Clustering Framework (OCF). The OCF groups the full set of spectral bands into $K$ clusters based on their statistical similarity. Following clustering, a Rank on Clusters Strategy (RCS) is applied to identify the most representative band within each cluster. This approach reduces redundancy and computational load while maintaining the integrity of the diagnostic spectral information necessary for distinguishing anomalies from the background.

2.2 Collaborative Representation (CR) Detection

The detection phase utilizes a dual concentric window approach. For each test pixel, a local background dictionary is constructed from pixels within an outer window, excluding those in an inner window to avoid target leakage. The center pixel is then modeled as a linear combination of these background neighbors using a distance-weighted regularization matrix.

The underlying logic is that background pixels can be accurately reconstructed by their neighbors, resulting in low residual errors. Conversely, anomalous pixels, which differ significantly from the background, will exhibit high reconstruction residuals. By performing this process on the optimized band subset, the contrast between background and anomaly is significantly amplified.

3. Experimental Results

The performance of BSCR was evaluated using three distinct hyperspectral datasets: the satellite-borne GF-5 (Gaofen-5) dataset, the airborne AVIRIS (San Diego) dataset, and the high-resolution Viareggio 2013 trial dataset.

3.1 Quantitative Performance

BSCR consistently outperformed baseline methods such as the classic Reed-Xiaoli (RX) detector and modern variants like PCA-based Collaborative Representation (PCAroCRD). On the GF-5 dataset, BSCR achieved a near-perfect AUC of 0.9946, compared to 0.4300 for the Subspace RX (SSRX) and 0.9933 for standard CRD. The "sweet spot" for the number of selected bands ($K$) was typically found to be between 20 and 45.

3.2 Qualitative Analysis

Separability maps (box plots of residual values) indicated that BSCR effectively suppresses background noise, pushing background residuals closer to zero while maintaining high values for anomalies. This leads to cleaner detection maps with significantly fewer false positives compared to methods that use the full spectral range or PCA-transformed data.

4. Discussion and Conclusion

The success of BSCR highlights the importance of preserving original spectral data over transformation-based dimensionality reduction in the context of anomaly detection. By selecting the most informative bands, the algorithm enhances the mathematical "predictability" of the background, making anomalies stand out more prominently. Future work will focus on automating the selection of optimal window sizes and improving computational efficiency for large-scale imagery.

Summary updated

Abstract

é«˜å ‰è°±é¥æ„Ÿå½±åƒå ·æœ‰å ‰è°±åˆ†è¾¨çŽ‡æžé«˜çš„ç‰¹ç‚¹ï¼Œæ‰¿è½½äº†å¤§é‡å¯åŒºåˆ†ä¸åŒç±»åž‹åœ°ç‰©çš„è¯Šæ–­æ€§å ‰è°±ä¿¡æ¯ä»¥åŠåŒºåˆ†äºšç±»ç›¸ä¼¼åœ°ç‰©ä¹‹é—´ç»†å¾®å·®åˆ«çš„å ‰è°±ä¿¡æ¯ï¼Œåœ¨ç›®æ ‡æŽ¢æµ‹é¢†åŸŸå ·æœ‰ç‹¬ç‰¹çš„ä¼˜åŠ¿ã€‚ä¸Žæ­¤åŒæ—¶ï¼Œé«˜å ‰è°±é¥æ„Ÿå½±åƒä¹Ÿå¸¦æ¥äº†æ•°æ®ç»´æ•°é«˜ã€é‚»è¿‘æ³¢æ®µä¹‹é—´å­˜åœ¨å¤§é‡å†—ä½™ä¿¡æ¯çš„é—®é¢˜ï¼Œé«˜ç»´åº¦çš„æ•°æ®ç»“æž„å¾€å¾€ä½¿å¾—é«˜å ‰è°±å½±åƒå¼‚å¸¸ç›®æ ‡ç±»å’ŒèƒŒæ™¯ç±»ä¹‹é—´çš„å¯åˆ†æ€§é™ä½Žã€‚ä¸ºäº†ç¼“è§£ä¸Šè¿°é—®é¢˜ï¼Œæœ¬æ–‡æå‡ºäº†ä¸€ç§åŸºäºŽæ³¢æ®µé€‰æ‹©çš„ååŒè¡¨è¾¾é«˜å ‰è°±å¼‚å¸¸æŽ¢æµ‹ç®—æ³•ã€‚é¦–å ˆï¼Œä½¿ç”¨æœ€ä¼˜èšç±»æ¡†æž¶å¯¹é«˜å ‰è°±æ³¢æ®µè¿›è¡Œé€‰æ‹©ï¼ŒèŽ·å¾—ä¸€ç»„æ³¢æ®µå­é›†æ¥è¡¨ç¤ºåŽŸæœ‰çš„å ¨éƒ¨æ³¢æ®µï¼Œä½¿å¾—é«˜å ‰è°±å½±åƒå¼‚å¸¸ç›®æ ‡ç±»ä¸ŽèƒŒæ™¯ç±»ä¹‹é—´çš„å¯åˆ†æ€§å¢žå¼ºã€‚ç„¶åŽä½¿ç”¨ååŒè¡¨è¾¾å¯¹å½±åƒä¸Šçš„åƒå ƒè¿›è¡Œé‡å»ºï¼Œç”±äºŽå¼‚å¸¸ç›®æ ‡ç±»å’ŒèƒŒæ™¯ç±»ä¹‹é—´çš„å¯åˆ†æ€§å¢žå¼ºï¼Œå¯¹å¼‚å¸¸ç›®æ ‡åƒå ƒè¿›è¡ŒååŒè¡¨è¾¾æ—¶å°†ä¼šå¾—åˆ°æ›´å¤§çš„æ®‹å·®ï¼Œå¼‚å¸¸ç›®æ ‡åƒå ƒçš„è¾“å‡ºå€¼å¢žå¤§ï¼Œå¯ä»¥æ›´å¥½åœ°å®žçŽ°å¼‚å¸¸ç›®æ ‡å’ŒèƒŒæ™¯ç±»çš„åˆ†ç¦»ã€‚æœ¬æ–‡ä½¿ç”¨äº†3ç»„é«˜å ‰è°±å½±åƒæ•°æ®è¿›è¡Œå¼‚å¸¸ç›®æ ‡æŽ¢æµ‹å®žéªŒï¼Œå®žéªŒç»“æžœè¡¨æ˜Žï¼Œè¯¥æ–¹æ³•ä¸Žå ¶ä»–çŽ°æœ‰é«˜å ‰è°±å¼‚å¸¸ç›®æ ‡æŽ¢æµ‹ç®—æ³•å¯¹æ¯”ï¼Œ 曲线下面积AUC(Area Under Curveï¼‰å€¼æ›´é«˜ï¼Œå¯ä»¥æ›´å¥½åœ°å®žçŽ°å¼‚å¸¸ç›®æ ‡ä¸ŽèƒŒæ™¯åˆ†ç¦»ï¼Œèƒ½å¤Ÿæ›´æœ‰æ•ˆåœ°å¯¹é«˜å ‰è°±å½±åƒè¿›è¡Œå¼‚å¸¸ç›®æ ‡æŽ¢æµ‹ã€‚
BibTeX
@inproceedings{zhu2020band,
  author = {Donglong Zhu and Bo Du and Liangpei Zhang},
  title = {{Band selection-based collaborative representation for anomaly detection in hyperspectral images}},
  booktitle = {National},
  year = {2020},
  doi = {10.11834/jrs.20209187},
}
Figure