Global Contextual Learning Network for Multisource Remote Sensing Image Fusion

Multisource remote sensing image fusion aims to combine the complementary advantage of different modalities for fine-grained classification. Current multisource image fusion methods are based on a patch-by-patch learning paradigm, which requires high computational burden. In addition, how to combine different attributes is still a challenging issue. In this article, a global contextual learning network (GCLNet) is studied for multisource remote sensing image fusion. In GCLNet, an encoder-decoder convolutional network is proposed to draw global spectral-spatial information, which can significantly reduce the computing cost. Furthermore, a contextual attention feature fusion module is designed to connect coarse-to-fine features between encoder and decoder. Experiments on Augsburg and Yellow River datasets reveal that the proposed GCLNet obtains promising results in the light of classification result and inference speed compared to other advanced fusion methods.