MCT-Net: A Multi-Scale Convolutional Transformer Network for Hyperspectral Unmixing

Hyperspectral unmixing (HU) has seen significant advancements through deep learning (DL) approaches, particularly autoencoder (AE) frameworks, due to their robust learning and data fitting capabilities. However, traditional convolutional autoencoders are constrained by local receptive fields, hindering connections between local spatial and global information. While Vision Transformers (ViT) excel in capturing long-range dependencies, current ViT-based methods often prioritize global relationships at the expense of fine-grained local features. Moreover, they underutilize spatial feature prior knowledge and require linear projection of each pixel for attention computation, leading to a large number of model parameters. To address these challenges, this paper proposes a Multi-Scale Convolutional Transformer Network (MCT-Net) for Hyperspectral Unmixing. MCT-Net integrates multi-scale convolutional autoencoders with a Transformer architecture enhanced by depth-wise separable convolutions. Following the autoencoder structure, our model consists of an encoder and a decoder. The encoder employs a multi-scale convolutional embedding module for comprehensive spatial feature capture, while the Transformer block incorporates depth-wise separable convolutions in place of traditional linear projections. This novel approach reduces model parameters, enhances focus on local spatial features, and improves computational efficiency. The decoder then reconstructs the hyperspectral image (HSI) from the learned representations, completing the autoencoder framework. Experimental results on the Samson dataset demonstrate MCT-Net's superior performance in both Root Mean Square Error (RMSE) and Spectral Angle Distance (SAD) metrics compared to traditional and DL-based unmixing methods.