SEFANet: Spectrum-Enhanced Feature Aggregation Network for Salient Object Detection in Hyperspectral Images
Kaihong Chen, Zhen Wang, Xuqi Wang, Shanwen Zhang · IEEE Access · 2026
Salient object detection in hyperspectral images (HSIs) is critical for applications in remote sensing, national defence, and environmental monitoring. However, the high dimensionality of HSIs, background interference, and scale variations pose significant challenges for traditional methods in effectively extracting spectral features and fusing spatial information. To address these issues, we propose a novel Spectrum-Enhanced Feature Aggregation Network (SEFANet) for hyperspectral salient object detection (HSOD). SEFANet adopts a dual-channel, dual-encoder architecture: the main channel leverages a High-Resolution Attention (HRA) module to model spatial and channel dependencies, while the auxiliary channel integrates a Feature Aggregation and Selection (FAS) module to progressively aggregate spectral features. Additionally, the Spectral Sensing Enhancement (SSE) module deeply fuses spatial and spectral information, enhancing robustness against background interference and scale variations. Extensive experiments on the HSOD-BIT dataset demonstrate that SEFANet achieves state-of-the-art performance, with an accuracy of 96.6%, mean intersection over union (mIoU) of 85.3%, mean Dice coefficient (mDice) of 88.1%, and a mean absolute error (MAE) of just 0.039. Compared to existing methods, SEFANet achieves significant improvements in both detection precision and robustness. The source code and dataset will be available on https://github.com/darkseid-arch/SEFANet.