Hyperspectral anomaly detection aims at identifying unique objects with different spatial and spectral appearances from its surrounding pixels. One of the widely used conventional methods is Collaborative-Representation-Based Detector (CRD). CRD approximates each pixel on the background that can be represented by neighboring pixels, while the anomaly pixels cannot be. The detection image generated from this method is quite satisfying since it can detect anomaly pixels quite accurately. However, the resulting detection images still contain a lot of pixels which are not anomaly even with tiny intensity values. In this study, we apply the Root-Mean (RM) adjustment threshold to filter the false-positive pixels in the output images of CRD, so that it gives more accurate results. In the experiment, six hyperspectral data with the corresponding ground truth images are used. The results show that after applying the RM threshold, the accuracy increases with the decrease of the six RMSEs in the range of 31.51 - 63.58%.
|Journal||IOP Conference Series: Materials Science and Engineering|
|Publication status||Published - 18 Mar 2020|
|Event||2nd International Conference on Engineering and Applied Sciences, InCEAS 2019 - Yogyakarta, Indonesia|
Duration: 16 Nov 2019 → …