Hyperspectral image classification has always been a task of great concern in the field of remote sensing. Transformer networks have demonstrated their powerful ability to capture global relationships and have achieved success in image analysis. However, when faced with the task of hyperspectral image classification, conventional image preprocessing methods lead to the loss of spectral information, failing to retain complete spectral-spatial information. To better solve this problem, this paper proposes an Attention-Guided CNN-Transformer Hybrid Network (AGCTHN). This network combines the strengths of Convolutional Neural Networks (CNNs) and Transformer networks in feature extraction while exploring the spectral and spatial information of large-scale and complex remote sensing images. A multi-level attention guidance mechanism is designed from shallow to deep at different scales, where convolutional features guide the self-attention module of the Transformer to focus on key details in the spectral bands of the image. Extensive experiments on hyperspectral image classification datasets validate the effectiveness of the proposed AGCTHN. The experimental results illustrate that the network outperforms state-of-the-art methods in classification accuracy. © 2023 IEEE.