Crop disease is a significant factor affecting crop yields, in severe cases, it can lead to a substantial reduction in crop production. However, most disease detection methods mainly rely on expert experience and general convolutional neural networks, but often face the problems of incorrect identification and difficulty in deploying resource-constrained environments. Therefore, this paper proposes a lightweight multilevel feature fusion classification model based on GhostNet (i.e., MLAF-G) to address the current challenges in crop disease identification. It improves performance through multilevel attention feature fusion. The backbone network combines an efficient computation (EC) module with the hybrid weight coordinate attention (HCA) module to improve feature map expressiveness, allowing for effective capture of crop disease characteristics with minimal computational cost. Additionally, an improved convolutional block attention module (I-CBAM) and a multilevel feature fusion mechanism are designed to integrate features across different levels, increasing accuracy and detection effectiveness. Extensive experiments on several datasets demonstrate that MLAF-G outperforms baseline models in classification accuracy, robustness,and efficiency. The experimental results prove that MLAF-G can efficiently accomplish crop disease detection tasks in resource-constrained environments, paving the way for wider application of intelligent agricultural technologies and contributing to the modernization and sustainable development of the agriculture industry.