REMOTE SENSING IMAGE CHANGE DETECTION BASED ON DEEP DICTIONARY LEARNING

As a hot topic in the field of remote sensing (RS), change detection aims to identify the semantic change between bitemporal RS images. Due to the semantic complexity of RS images, how to accurately detect the semantic change has become a challenging problem. Recently, many deep-based methods are proposed to solve this issue. However, ignoring the representation difference of same semantics in different periods limits their performance, such as river is liquid in summer and solid in winter. Therefore, a new method is presented, named dictionary learning based change detector (DLCDet), which consists of feature pyramid network, deep dictionary learning and dual supervision modules. In DLCDet, the deep dictionary learning is proposed to reduce the representation difference so that DLCDet identifies the potential semantic change more accurately. Experiments are conducted on two public datasets change detection dataset (CDD) and building change detection dataset (BCDD), which demonstrates the effectiveness of the proposed method.