An Unsupervised Deep Hyperspectral Anomaly Detector.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral imaging (HSI) offers object discrimination via spectral signatures.
• Traditional anomaly detection in HSI faces challenges with background estimation and false alarms.
• Existing methods like Reed-Xiaoli (RXD) and Collaborative Representation Detector (CRD) have limitations.

Purpose of the Study:

• To propose a Deep Belief Network (DBN) based anomaly detector for HSI.
• To enhance detection accuracy by learning high-level features and reconstruction errors.
• To mitigate false alarms caused by local anomalies.

Main Methods:

• A Deep Belief Network (DBN) is employed for unsupervised feature learning and anomaly detection.
• Adaptive weights are derived from reconstruction errors and statistical information to reduce anomaly contamination.
• A modified code image from DBN inference and local Euclidean distance are used for target identification.

Main Results:

• The proposed DBN-based detector demonstrates superior performance on synthetic and real HSI datasets.
• It effectively learns features and reconstruction errors, independent of background assumptions.
• Adaptive weighting successfully reduces contamination from local anomalies.

Conclusions:

• The DBN-based anomaly detector offers a robust and accurate solution for HSI analysis.
• This approach overcomes limitations of traditional and state-of-the-art detectors.
• It shows significant potential for applications in agriculture, environmental monitoring, and defense.