Anomaly Detection Based on Sigmoid Metric and Object Area Filtering in Hyperspectral Images

This paper outlines a new approach to detect anomalies in hyperspectral images based on peripheral pixels. The proposed methodology contains two main steps. First, a new distance score is introduced based on the sigmoid function and root mean square error (RMSE). We estimate how likely the target pixel is an anomaly by averaging the new metric over its neighboring window.

Second, a state-of-the-art method is applied to eliminate unacceptable objects according to their size. In this light, the objects whose size is out of an acceptable interval are removed.

Comprehensive experimental evaluations have been conducted to confirm that the proposed method significantly outperforms several recent algorithms in accuracy and computational time.