Open set classification of Hyperspectral images with energy models

Hyperspectral images are rich in spectral data, lending themselves well to pixel-level image classification tasks. Previous studies primarily focus on closed-set classification within the realm of hyperspectral image classification. However, real-world scenarios present the challenge of dealing with object classes not encountered during the training phase, a scenario known as open set classification, which has garnered less attention compared to the closed set paradigm. In this paper, we propose a methodology anchored on ConvMixer for tackling open-set classification by utilizing energy-based models. We incorporate a Selective Kernel Attention (SKA) to capture the notion that different feature maps usually correspond to different objects in deep neural networks. Our experimental validation, conducted on two datasets, specifically the WHU-Hi-HanChuan and WHU-Hi-HongHu datasets, showcases promising outcomes of the introduced method.