Self-learing Driven For Semantic Segmentation Of Remote Sensing Images Based On Random Field And CNN

The continuous development of remote sensing technology enables people to obtain remote sensing images with higher spatial resolution,more wavebands,and shorter return visit periods.Making full use of the information in remote sensing images and accurately classify it has great significance in many fields such as precision agriculture,target recognition,disaster monitoring.In recent years,convolutional neural networks(CNN)have been widely used in semantic segmentation of remote sensing images,and it have achieved great success.However,in heterogeneous Earth surface systems,the characteristics of similar ground features often have regional characteristics,which can change when the region changes,and some ground features can also change with time.It makes the CNN’s application performance in large regions poor.And CNN is a data driven model which lacks sufficient mathematical theory to support,so it also has the problem of poor model interpretability.In order to break through the bottleneck faced by CNN,geoscientific knowledge information is used to assist CNN models in semantic inference.Geographical knowledge inference refers to the establishment of new relationships between ground objects and objects through computer inference based on the relationship between the entity concepts of ground objects in the map of geoscience knowledge,and the discovery of new geoscience knowledge.Using geographical knowledge for auxiliary inference can compensate for the weakness of the poor interpretability of CNN models and enhance the application performance of CNN in large regions.Currently,statistical reasoning is a commonly used knowledge reasoning method,while Markov Random Field(MRF)is a probability graph model with a solid theoretical foundation in statistics.It describes spatial context relationships in images from a statistical perspective,uses likelihood functions and joint probabilities to describe data feature information and geographical knowledge reasoning,and optimizes semantic segmentation results by maximizing a posterior probability based on Bayesian theorem.Therefore,this paper considers using MRF to jointly model data driven CNN methods and geographical knowledge in remote sensing images,and designs two data knowledge driven semantic segmentation algorithms that cooperate with random fields and CNN,optimizing the semantic segmentation results.The research in this article is as follows:Due to the diversity of spatial distribution of land objects in remote sensing images,it is difficult to effectively learn general geographical rules and apply them to specific images.In order to utilize geographical knowledge more effectively,the first work of this paper proposes a random field and CNN collaborative model(SLU-CNN).This model learns specific spatial dependencies between different objects based on the prediction results of CNN,and combines them with the results of CNN using MRF to perform semantic inference through statistical inference.SLU-CNN mainly involves two modules.Firstly,a mean shift algorithm is used to generate inference units from the prediction results provided by CNN.Secondly,selflearning is used to establish specific adaptive geographic relationships based on spatial dependencies between inference units.Finally,a logarithmic function is selected to transform this geographical relationship into an adaptive anisotropic matrix,which is embedded in an object based MRF model.By using statistical inference,collaboration between CNN results and semantic inference is achieved.Experiments were conducted on the dataset of GF-2 and Sentinel-2,and compared with different CNN methods to verify the effectiveness of this method.As the spatial resolution of remote sensing images becomes higher and higher,the semantic information contained in them becomes more and more abundant.Using only one layer of semantic information is difficult to meet the requirements of segmentation tasks.Therefore,the second work of this article extends the previous model and designs an algorithm(SLU-MLS-CNN)that can process remote sensing images containing two-level semantic information.This model establishes a hybrid tag field,models two-level semantics in the form of vectors,and designs a new joint distribution to capture the anisotropic spatial interactions within the same semantic layer of the hybrid tag field,as well as the anisotropic spatial interactions between different semantic layers.The interaction between high-level and low-level semantic information is iteratively optimized,ultimately providing a better prediction result for both semantic layers.This method is tested on the GF-2 dataset with 5 and 15 semantic layers,and the results verify the effectiveness of this method.