SRGAN-ConvFormer: A Dual-Channel transformer framework for hyperspectral image classification

Spatial resolution is crucial for accurate hyperspectral image classification. Low spatial resolution is often responsible for mixed pixels in an image that reduces discriminability between land-cover classes. In this work, the resolution of hyperspectral images is improved by using super-resolution generative adversarial network (SRGAN) that creates sharper images by improving edges and enhancing texture details. A new SRGAN-ConFormer is proposed that integrates SRGAN and ConvFormer-based transformer models using a dual channel convolutional neural network (DCCNN) into a single classification framework. The SRGAN model generates enhanced spatial features, while spectral information is processed by ConvFormer transformer model. The high resolution spatial patches and spectral features are integrated by DCCNN block. SRGAN-ConvFormer allows effective exploitation and adaptive fusion of complementary spectral and spatial information. The performance of the proposed SRGAN-ConvFormer evaluated on four benchmark datasets: Botswana, Pavia University (PU), Kennedy Space Center (KSC) and Salinas to assess the performance of the proposed method. The experimental results demonstrate the remarkable performance of the proposed method with overall accuracies of 99.26% (κ = 99.19%), 99.78% (κ = 99.76%), 99.86% (κ = 99.81%) and 99.75% (κ = 99.73%) for Botswana. KSC, PU, and Salinas respectively. The results exhibit the robustness and generalization in both coarse-resolution datasets such as Botswana and KSC, as well as fine-resolution datasets PU, and Salinas. The results also demonstrated that the proposed SRGAN-ConvFormer method outperformed several state of the art methods.