Joint Direct and Cascaded Channel Estimation Using TCNNs in XL-RIS-Aided Massive MIMO

Reconfigurable Intelligent Surfaces (RIS) enable energy-efficient wireless control via passive reflection. As RIS scale to Extremely Large RIS (XL-RIS) with ultra-dense elements, Channel State Information (CSI) estimation becomes more challenging due to high dimensionality and structured sparsity. This paper first introduces a dual-branch pooling module into the Twin Convolutional Neural Network (TCNN) architecture to jointly estimate direct and cascaded channels. The enhanced TCNN embeds this hybrid pooling-combining max pooling for local feature capture and attention pooling for global context modeling-within each convolutional stream. This design improves the network’s ability to exploit the sparsity and structured features of XL-RIS channels. Extensive simulations verify the effectiveness and scalability of the proposed method, which outperforms existing deep learning-based estimators in accuracy and robustness while reducing pilot overhead in largescale XL-RIS deployments.