Enhancing Confidence of Unsupervised Change Detection in Hyperspectral Images

Change detection in bi-temporal hyperspectral images is important for many environmental applications. A popular transfer learning-based approach for change detection in multispectral images is to use a pre-trained network as a feature extractor. However, such pre-trained networks are not generally available in the case of hyperspectral images. To circumnavigate the requirement of the pre-trained network, recent works used untrained models as a bi-temporal feature extractor. However, results obtained using such an untrained model can be of varying reliability. We postulate that different models obtained with slight perturbation in the untrained model can yield an ensemble of change detection maps. The variance computed over such an ensemble can be used as an indicator of confidence or certainty. Our experiments on two datasets show that the proposed strategy can identify pixels in an analyzed scene where the result is more certain.