Harmonising Topology and Labels: An Adaptive Graph Learning Framework for Hyperspectral Image Classification

Hyperspectral image (HSI) classification is a critical task in remote sensing with applications spanning urban planning and environmental monitoring. The inherent challenge lies in the spectral-material inconsistency across heterogeneous regions, which complicates the modelling of spatial-spectral information. While Graph Convolutional Networks (GCNs) offer a powerful paradigm for this purpose, they are prone to over-smoothing and often lack an adaptive mechanism to discriminate node importance, particularly in spectrally mixed regions. Moreover, existing graph-based methods underutilise sparse label information. To address these issues, we propose an adaptive graph learning framework that introduces a novel Topology-Label Scoring (TLS) mechanism to identify and prioritise nodes in spectrally mixed regions. This scoring function integrates topological uniqueness with neighbourhood label consistency, enabling sensitive perception of heterogeneous zones. Subsequently, a Mixed-Superpixel CNN (MSC) module extracts fine-grained, local features exclusively from these high-scoring, mixed nodes. The framework culminates in a fusion of superpixel-level global features from the GCN with pixel-level local features from the MSC. Extensive experiments on four benchmark HSI datasets demonstrate that our method consistently achieves superior classification accuracy, significantly outperforming state-of-the-art techniques. This work not only provides a robust solution for HSI classification but also introduces a principled approach for adaptive feature fusion in graph-based learning, paving the way for more nuanced and efficient analyses of complex, high-dimensional data. The source code will be freely available at https://github.com/jieqian0816/TLS-MSC