MLRSNet: capsule networks-based multi-label hyperspectral image retrieval

With the growing volume and complexity of remote sensing data, effective multi-label image retrieval has become a critical challenge in hyperspectral image analysis. While deep learning techniques have significantly enhanced classification performance, retrieval models that preserve spatial hierarchies and semantic co-occurrence remain underexplored. In the present study, we propose MLRSNet, a capsule network-based architecture tailored for multi-label hyperspectral image retrieval. MLRSNet integrates capsule routing and a reconstruction loss mechanism to maintain part-whole relationships and improve discriminative feature learning in complex scenes. For evaluation, we employed two benchmark datasets, a publicly available relabeled version of the UC Merced Land Use dataset, constructed using a semi-supervised graph-theoretic method in prior work, and the Ankara hyperspectral imaging (HSI) dataset, which contains native multi-label annotations. These datasets enabled the evaluation of retrieval performance in realistic multi-label scenarios. Experimental results across five random train-test splits show that MLRSNet consistently outperforms conventional CNN-based approaches in terms of accuracy, precision and recall.