Title: Transformer-Guided Content-Adaptive Graph Learning for Hyperspectral Unmixing

Abstract:

Hyperspectral unmixing (HU) aims to decompose mixed pixels found in remote sensing imagery into constituent endmembers alongside their respective abundance levels. While deep learning has driven substantial advancements in this domain, many existing approaches struggle to concurrently model global dependencies and local consistency. This limitation hinders the ability to maintain both long-range interactions and fine-grained boundary details. To address these issues, this study introduces a novel framework known as Transformer-Guided Content-Adaptive Graph Unmixing (T-CAGU). The proposed method utilizes a transformer to capture global dependencies and incorporates a content-adaptive graph neural network to strengthen local relationships. In contrast to prior works, T-CAGU employs multiple propagation orders to dynamically learn the graph structure, thereby ensuring robustness against noise. Additionally, the framework adopts a graph residual mechanism designed to preserve global information and stabilize the training process. Experimental evaluations confirm that T-CAGU outperforms current state-of-the-art methods. The source code is publicly accessible at https://github.com/xianchaoxiu/T-CAGU.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC