Learning a Cross-Modality Anomaly Detector for Remote Sensing Imagery.

Area of Science:

• Earth Observation
• Computer Vision
• Machine Learning

Background:

• Remote sensing anomaly detection is crucial for Earth monitoring.
• Current models struggle with cross-modality transfer due to background-specific training.
• A flexible, cost-effective solution is needed for diverse anomaly types and data sources.

Purpose of the Study:

• To develop a transferable remote sensing anomaly detection model.
• To enable cross-modality detection across hyperspectral, visible, SAR, infrared, and low-light imagery.
• To overcome limitations of background-dependent anomaly detection methods.

Main Methods:

• Proposed a learning target conversion from background distribution to a consistent deviation metric.
• Developed two large-margin losses for pixel-level and feature-level deviation ranking.
• Utilized anomaly simulation strategies for model training due to scarcity of real anomalies.

Main Results:

• The proposed method achieves cross-modality detection in a zero-shot manner.
• Demonstrated successful anomaly detection across five different sensor modalities.
• The large-margin learning ensures the transferability of the learned deviation metric.

Conclusions:

• The developed model offers a cost-effective and flexible approach for Earth monitoring anomaly detection.
• The large-margin learning strategy is key to achieving cross-modality and zero-shot detection capabilities.
• This work advances the field of remote sensing by enabling robust anomaly detection across diverse data sources.