MGMNet: Mutual-Guidance Mechanism for Joint Classification of Multisource Remote Sensing Data

The joint classification of multisource remote sensing data has shown significant potential in the precise interpretation of land cover. Existing methods mainly employ a dual-stream architecture to independently extract features, subsequently merging them through a fusion technique. However, the majority of these methods concentrate solely on extracting features and overlooking the importance of feature interaction. In fact, different source data of the same scene exhibit strong complementary information, which is useful for excavating potential information. To this end, we propose a mutual-guidance mechanism (MGM) in this article to explore the information interaction between multisource data. Specifically, we first embed the MGM into a two-stream network, called MGMNet, where the image features extracted from one source data are convolved with context-aware filters generated from another source data. Through such an interaction way, each source data provide the rich guidance information for another source data during the feature extraction process. Subsequently, the extracted multisource features are fused with an attentional feature fusion module, resulting in a highly discriminative integrated feature for the classification task. At last, a weighted decision-level fusion module is employed to enhance the precision of the classification. To validate the effectiveness of MGMNet, comprehensive experiments are carried out on three multisource datasets. The conclusive results from these experiments indicate that MGMNet surpasses some competitive approaches.