Knowledge graph-guided deep network for high-resolution remote sensing image scene classification
Thanks to the rapid development of deep network theory and methods,deep networks have gradually become the mainstream technology for remote sensing image scene classification tasks.However,existing deep network-based remote sensing image scene classification methods are highly dependent on a large number of manually labeled training samples and cannot effectively integrate and utilize the rich prior knowledge in the remote sensing field.In order to improve the utilization of domain knowledge while reducing the dependence on labeled samples,this paper proposes a knowledge graph-guided deep net-work learning method for high-resolution remote sensing image scene classification.First,this paper constructs a land cover concept knowledge graph that includes various sources of knowledge in the field to more flexibly and conveniently apply domain prior knowledge.Furthermore,through the knowledge graph representation learning method,the semantic categories of re-mote sensing scenes in the land cover concept knowledge graph are expressed as semantic vectors to form a semantic benchmark for remote sensing scene categories.In the knowledge-guided learning stage,the cross-modal alignment constraint between the scene category semantic vector and the shallow visual feature vector of the deep network is applied to guide the shallow part of the deep network to more effectively learn shared features of different categories of remote sensing image scenes,while in the deep part of the deep network,it is still guided by scene category labels to learn discriminative features of different remote sensing scenes.In the testing stage,the optimized deep network model can complete high-precision remote sensing image scene classification without relying on any prior knowledge.The experimental results on the currently largest publicly available re-mote sensing image scene classification dataset show that the proposed knowledge-guided learning method can obtain optimal classification performance at different training sample ratios such as 10％,30％,and 50％compared with existing methods.Under the condition of 10％sample ratio,our proposed method can achieve an improvement of 5.11％in overall accuracy(OA)compared with baseline deep networks.
remote sensing image scene classificationland cover concept knowledge graphknowledge graph representation learningknowledge graph-guided deep network
万方数据
维普
