Change Detection in Hyperdimensional Images Using Untrained Models

Deep transfer-learning based change detection methods are dependent on the availability of sensor-specific pre-trained feature extractors. Such extractors not always available due to lack training data, especially for hyperspectral sensors and other hyperdimensional images. Moreover models trained easily multispectral (RGB/RGB-NIR) images cannot be reused such their irregular number bands. While show large spectral bands, they generally much less spatial complexity, thus reducing requirement receptive fields convolution filters. Recent works in computer vision have shown that even untrained deep can yield remarkable result some tasks like super-resolution surface reconstruction. This motivates us make a bold proposition lightweight model, initialized with weight initialization strategy, used extract useful semantic features from bi-temporal Based this proposition, we design novel framework by extracting using an model further comparing extracted Change Vector Analysis distinguish changed pixels unchanged ones. We use hypervectors cluster into different groups. conduct experiments four datasets: three datasets Polarimetric Synthetic Aperture Radar dataset. The results clearly demonstrate proposed method is suitable remote sensing data.