Open-Set Hyperspectral Land-Cover Classification Using Second-Order Attention and Gaussian Mixture Clustering

Open-set hyperspectral classification requires accurate recognition of known land-cover classes and detection of samples from unseen classes. We propose GRDOSR, which extends representative-discriminative open-set recognition (RDOSR) by combining second-order attention with Gaussian mixture modeling. Spatial and channel attention capture local feature dependencies, while mixture components define specialized networks that jointly learn reconstruction and classification. Their predictions are combined using mixture posteriors, and a score based on reconstruction error and mixture likelihood is calibrated using known validation data for unknown rejection. Across five hyperspectral benchmarks, GRDOSR improves the mean area under the receiver operating characteristic curve (AUC) by 1.43–11.57 percentage points and known-class overall accuracy (OA) by 0.27–13.57 points relative to RDOSR, with a higher mean open-set classification rate (OSCR) on every dataset. It achieves the highest mean AUC and OSCR on Indian Pines and Pavia University datasets among the compared methods. These results support the use of local context and modeling of known feature distributions for land-cover mapping when the training data cover only part of the class inventory.