Superpixel Prior Cluster-Level Contrastive Clustering Network for Large-Scale Urban Hyperspectral Images and Vehicle Detection

How to get effective sample representations when the training labels are absent is always a problem for hyperspectral image clustering. Some contrastive clustering models guide the representation learning process to adapt to clustering task but these models are still instance-level, which cannot match the clustering task. This article proposes a Superpixel Prior Cluster-level Contrastive (SPCC) clustering network for large-scale hyperspectral image clustering tasks. Instead of instance-level contrastive learning, SPCC retains the useful representations selectively under the guidance of the superpixel subregion. The proposed framework treats each sample as a view for its belonged cluster and learns cluster-level semantic representations. SPCC enhances the separability of samples by employing contrastive loss, which proves effective for vehicle detection tasks. Extensive experiments are designed to obtain competitive results for the popular clustering methods on three Hyperspectral image datasets. The experimental results indicate that SPCC performs well in both hyperspectral image clustering and vehicle recognition tasks.