Dual partial recurrent networks for hyperspectral image change detection

This paper presents a Dual Partial Recurrent Networks (DUAL-PRNs) which can project more accurate and effective image features by learning invariant pixel pairs with high confidence. The Change Vector Analysis provides a reference for the model to select invariant pixel pairs with high confidence as training samples. Then, the Unsupervised Slow Feature Analysis (USFA) is utilized to suppress the invariant pixel features projected by DUAL-PRNs, and highlight the variant pixel features, respectively. Thus, more obvious discrimination between the invariant and variant pixels can be achieved. Two groups of features are then obtained by passing bi-temporal remote sensing images through DUAL-PRNs and USFA. Chi-square distance is employed to calculate the divergence between two groups of features and thus generate the Change Intensity Map. Finally, the thresholding algorithm transforms the change intensity map into binary change map. Experimental results show that the proposed change detection model DUAL-PRNs performs better than the advanced model DSFA-128-2.