Timely and accurate identification of tomato leaf disease types can effectively improve the quality and yield of tomatoes, increase farmers' economic returns, and promote intelligent and modernized tomato production. To address the problems of intra- and inter-class multi-scale variation, complex background interference, and difficulty in mobile model deployment faced by tomato leaf disease identification, we propose a lightweight Ghost Dense network (LGDNet) to identify diseases of tomato leaves. First, we replace the standard convolution of the bottleneck layer in DenseNet with the Ghost module, which compresses the network size while maintaining the model's adaptability to the multi-scale variation of tomato leaf diseases. Then, we propose a lightweight and efficient coordinate multidimensional information fusion attention (CMIFA) module that enhances feature extraction for tomato leaves and enables the model to locate the diseased areas more accurately. The experimental results indicate that LGDNet reaches optimal recognition performance in both the PlantVillage dataset with a simple background and the Dataset of Tomato Leaves with natural scenes. Moreover, LGDNet achieves the minimum number of parameters among the compared models. In summary, LGDNet provides an excellent solution to the problem of accurately identifying tomato leaves in complex environments and provides a reference for deployment on mobile.