Open Set Land Cover Classification Based On Embedded Space Feature Analysis

For a long time,people have done a lot of research on land cover classification,not only because it is a hot topic of global environmental research,but also because the most important thing is that the study of land cover classification is of great significance to global socio-economic and ecological environment.In the current land cover classification research,we used to assume that all categories are known,that is,analyze the land cover type in a relatively closed setting.Therefore,once an unknown category sample appears in the network,the system will force it to be divided into a known category,thus affecting the classification accuracy of the system.In the representative-discriminative open set recognition(RDOSR)framework,the open set land cover classification problem is mainly studied.It is required that not only the samples belonging to unknown categories can be identified during the test process,but also the performance of known categories can be maintained.The framework is mainly composed of two parts:1.The closed set embedding component projects the data from the original image domain into the embedding domain,making it easier to distinguish different categories with similar spectral characteristics;2.The multi-task representation discrimination learning component learns the finer scale representation scheme in the abundance space,so as to better distinguish the unknown category from the known category.The weights of closed set embedded components are trained with known category data.The multi-task representation discrimination learning component improves the representativeness and discrimination ability of the extracted feature vectors,thus making the network more informative and effective in identifying unknown categories.The component is composed of coder-decoder architecture,in which the representative features s are extracted by sparse Dirichlet encoder E,and the decoder is formed by the basis of known class training.In addition,the component also includes a classifier C applied to representative features s to further improve the identification ability.In this way,when the unknown category is input into the network,it can generate higher reconstruction error,so that the unknown category can be detected.Finally,the effectiveness of this method is proved on multiple data sets.The research contents of this paper are as follows:1.The open set land cover classification based on clustering analysis is studied.In the RDOSR framework,the embedded spatial featuresz_F are reconstructed using an encoder decoder architecture,and unknown categories are detected by the magnitude of the reconstruction error.However,the algorithm does not consider the problem that the reconstruction error of known categories may be large or that the reconstruction error of known and unknown categories is similar.This paper proposes to map the image from the original spatial domain to the embedded spatial feature domain during the training process,extract the embedded features of known categories,perform clustering analysis on their embedded features,and then generate multiple clusters.From the perspective of data or features,the similarity between known categories in each cluster is higher,and then use the RDOSR framework to conduct separate training for each cluster to reduce the reconstruction error of the training data for known categories,The reconstruction error of unknown class training data is relatively larger.During the testing process,for the identified samples,first map them to the embedding space,then determine which cluster they belong to based on the clustering parameters,and then apply the RDOSR framework corresponding to the cluster parameters for classification.If the final reconstruction error is greater than the set threshold,it is determined as an unknown category.Finally,experiments show that the open set land cover classification algorithm based on clustering analysis designed in this paper can effectively identify unknown classes,maintain the performance of known classes,and improve classification accuracy.2.Aiming at the representative-discriminatory open set recognition(RDOSR)framework,which refers to a classifier F as a network structure for embedding a closed set into a feature space,this paper designs a multi embedded feature fusion network using the structure of the Transformer model’s multi head attention mechanism for reference.It detects unknown classes in the multi embedded feature fusion spatial domain projected by the closed set embedding layer,rather than in the image domain.By connecting multiple embedded feature extraction modules in parallel,Multiple classifiers_tF are spliced to extract embedded feature vectors,and then multiple embedded features are integrated in a certain way to obtain more informative fusion embedded features.While ensuring the accuracy of the closed set embedded learning network,the extraction of embedded features is further improved.Finally,the encoder decoder architecture is used for reconstruction,and the magnitude of the reconstruction error is used to detect unknown categories,resulting in better classification performance,It better solves the problem that different types of satellites may have similar spectral characteristics,which is beneficial for the network to distinguish unknown classes that are close to known class spectra.Experiments show that the multi embedded feature fusion mechanism can achieve good classification results in open set land cover classification.