Spatial-Spectral Correlation Network With Super-Resolution Learning for Hyperspectral Target Detection

Hyperspectral target detection (HTD) aims at extracting targets from complex backgrounds while overcoming noise interference. Existing deep learning models for HTD usually suffer from low spatial resolution and unitary representation, especially in spaceborne platforms. Super-resolution, as a critical technology to enhance the spatial details, could effectively address the aforementioned issue. To improve target detection through super-resolution, we propose an end-to-end spectral-spatial correlation network (SR-HTD) with joint spatial and spectral regularization for high-precision hyperspectral target detection. Specifically, a spatial correlation aggregation module inspired by latent low-rank representation is designed to capture global spatial structures absent in real scenarios. This enables super-resolution (SR)-HTD to reason about the underlying target scene and perform background estimation. To further enhance spectral-spatial regularization, the SR branches are integrated into the generative reconstruction module to emphasize the structural and spectral differences between target and background classes. In contrast to conventional models, the cascaded and complementary mechanisms fully benefit from HTD and SR, forming a more comprehensive solution and releasing barriers to target detection from a complex and noisy background. We generated synthetic panchromatic and hyperspectral images (HSIs)to evaluate the target detection performance on high-resolution images and the impact of SR on target detection. For most of the datasets tested, the detection accuracy of AUC(D,F) is the best and second best among all the methods compared. Compared with state-of-the-art methods, the proposed method shows the expected superiority of the airport beach urban and hyperspectral digital imagery collection experiment datasets.