The Methodology Research On Geographic Knowledge-Guided Deep Semantic Segmentation Of Remote Sensing Imagery

As the fundamental task of geographic information interpretation,semantic segmentation of remote sensing(RS)imagery,which has important application value,is widely used in the fields of land cover mapping,urban spatial planning,and environmental disaster monitoring.The era of high-resolution earth observation has arrived,remote sensing images contain more and more information.The complex coverage of objects in the image has led to the serious phenomena of intra-class variability and inter-class similarity,which makes the RS image semantic segmentation full of challenges.However,traditional semantic segmentation methods rely on artificially designed feature descriptors.It is difficult to complete high-precision segmentation of high-resolution remote sensing images,which cannot meet the needs of practical applications and research.With the rapid development of artificial intelligence,deep learning methods are gradually used in remote sensing image interpretation.Compared with traditional methods,the deep semantic segmentation network,which is one of the representatives of deep learning,has made significant progress in the semantic segmentation of remote sensing images.Although deep learning methods have powerful learning capabilities for deep features,deep learning is a data-driven method.Its learning process is highly dependent on reducing pixellevel losses to reversely optimize network parameters.As it is difficult to use rich highlevel knowledge(including prior knowledge and semantic information),the deep semantic segmentation network lacks the ability of self-recognition with the poor reliability and interpretability of its output.However,the prior knowledge and the semantic information are essential for the high-precision and interpretable interpretation of remote sensing data,just as the human brain advanced intelligent system is based on knowledge and semantics to reason,so as to make the high reliability and explainable classification decision.Although knowledge-driven methods are highly explainable,these methods have limited accuracy due to the technical challenges of complete knowledge modeling.In view of the precision advantages and powerful learning capabilities of data-driven deep learning methods,the construction of the intelligent interpretation method for remote sensing images embedded with geographic knowledge is the key to give full play to the advantages of data-driven and knowledge-driven methods.It not only takes the advantage of deep learning of being good at learning low-level features which are difficult to accurately express,but also combines the advantages of high interpretability and reliability of the knowledge reasoning for classification.According to the above analysis,this paper proposes the geographic knowledgeguided deep semantic segmentation method of remote sensing imagery where geographic knowledge is embedded in the deep semantic segmentation network at the three levels of external assistance,internal participation and autonomous learning to guide the training of the network,thereby improving the reliability and interpretability of the segmentation.The research content and innovations of this paper are as follows:(1)At the level of external assistance,aiming at the problem that deep neural networks lack the use of geographic knowledge,a deep semantic segmentation method guided by geographic knowledge reasoning is proposed,which realizes the coupling of deep semantic segmentation network and knowledge reasoning.Knowledge reasoning uses high-level geographic knowledge to guide the interpretation,directly corrects the misclassification of the deep semantic segmentation network,and automatically extracts additional information to indirectly assist the neural network.The whole process forms a loop,which is continuously optimized through iteration.(2)At the level of internal participation,the deep semantic segmentation method driven by graph convolutional neural network is designed to solve the problem of deep neural network's ignorance of the spatial semantic relationship between objects,which introduces the spatial semantic information into the deep semantic segmentation.In the method,the deep semantic segmentation network extracts deep features from the image to initialize the graph nodes,and the graph convolutional neural network uses the powerful modeling ability of the dependency relationship between the graph nodes to introduce the spatial relationship between the objects into the classification.(3)At the level of autonomous learning,this paper constructs the deep semantic segmentation method guided by remote sensing knowledge graph to realize the autonomous learning of object-level geographic knowledge by the deep semantic segmentation network.The method uses the geographic knowledge extracted from the remote sensing knowledge graph to construct loss constraints,including object-level connectivity constraints and inter-object coherence constraints,which autonomously guides the learning process of deep semantic segmentation network.Geographic knowledge is a key element in constructing a new generation of remote sensing big data artificial intelligence interpretation theory and method with selfcognition ability.In order to improve the intelligent recognition of deep semantic segmentation network,this paper designs the geographic knowledge-guided deep semantic segmentation method of remote sensing imagery,which realizes the coupling of knowledge-driven symbolic reasoning and data-driven deep learning at the levels of external assistance,internal participation and autonomous learning.Therefore,a new approach is explored for the deep semantic segmentation method guided by geographic knowledge.Deeper integration of knowledge reasoning and deep learning to autonomous learn the reasoning process will be the important research in the future.The experimental results show that the proposed method greatly enhances the performance and robustness of the deep semantic segmentation network through the embedding and reasoning of geographic knowledge.At the same time,benefiting from the knowledge reasoning,it effectively improves the reliability and interpretability of the segmentation.