<p>Hyperspectral image (HSI) classification is a fundamental remote sensing task, and deep learning has become the mainstream paradigm for it. However, existing deep learning models for HSI classification suffer from three critical limitations: insufficient local feature extraction, severe spectral-spatial feature coupling, and redundant token interference, which drastically restrict classification performance. To overcome these core limitations, this paper proposes a spectral-spatial decoupled attention network (SSDAnet) for HSI classification, which realizes decoupled modeling of spectral-spatial features and adaptive suppression of redundant tokens through a collaborative convolutional-Transformer architecture. Specifically, we design a spectral-spatial decoupled convolution module (SSDCM) to extract robust spectral-spatial features by integrating pixel-adaptive calibration and residual connections, addressing weak local feature perception and feature coupling. We then develop a tokenized feature construction module (TFCM) to convert high-dimensional features into compact, discriminative token representations, laying a foundation for efficient attention modeling. Finally, we propose a spectral-spatial dual attention module (SSDAM), which uses channel-wise adaptive gating to enhance local feature expression, adopts dual-channel decoupled attention to thoroughly disentangle spectral-spatial features, and utilizes Gaussian-guided soft-mask token selection to eliminate redundant token interference, fundamentally solving the inherent drawbacks of existing Transformers. Experimental results on the Pavia University, Indian Pines, Salinas, and WHU-Hi-LongKou datasets demonstrate that the proposed method achieves overall accuracies of 92.03%, 83.69%, 96.24%, and 95.57% with only 10 training samples per class, benefiting from spectral-spatial decoupling and redundant token suppression, thereby showing superior classification performance.</p>