<p>Hyperspectral image (HSI) classification methods based on diffusion models have made significant progress with limited labeled samples. Existing Transformer-based encoder-decoder architectures with multi-head attention has demonstrated excellent performance in capturing complex dependencies. However, it often leads to the waste of computing resources and reduction in computational efficiency in the application of diffusion model. Additionally, traditional fine-tuning strategies rely on large amounts of labeled data to optimize model performance, which can result in overfitting and negatively impact classification ability of the model. To overcome these limitations, a lightweight HSI classification method-single-head attention-driven diffusion masked autoencoder (SHAD-MAE) is proposed. SHAD-MAE adopts a single-head attention mechanism during the processes of encoder and decoder to save resources and enable the model to learn more effective features. Moreover, we introduce a novel contrastive learning method to enhance the ability of the model to perceive subtle feature changes in HSI data. Finally, SHAD-MAE utilizes an asymmetric encoder-decoder architecture with a long skip connection in the encoder, which reduces information degradation. Results on four HSI datasets demonstrate that SHAD-MAE outperforms existing methods in classification accuracy and achieves improvements in computational efficiency.</p>