Abstract: To intelligently identify the growth and development status of tobacco plants in fields, a gradient fertilization field experiment was conducted and an image dataset containing 6 029 images of tobacco plants with different nutritional status throughout the growth and development cycles was generated using a UAV data acquisition platform. Based on the classical image classification network EfficientNet-B0, a lightweight ES-EfficientNet model for rapid identification of tobacco growth stages and nutritional status was developed by replacing the original squeeze-and-excitation (SE) attention mechanism in MBConv with an efficient channel attention (ECA) mechanism, and further by replacing standard convolutions with grouped convolutions, and incorporating a channel shuffle strategy to reduce model parameters. The results showed that: 1) Compared to the original EfficientNet-B0 model, the recognition accuracy of the ES-EfficientNet model increased by 2.4 percentage points and the number of parameters decreased by 2.26 MB. 2) Compared to common image classification models such as EfficientNet V2, Mobile ViT, MobileNet, ResNet and ShuffleNet, the recognition accuracy of the developed model increased by 0.7 to 43.8 percentage points. Considering both accuracy and number of parameters, the ES-EfficientNet model has a reduced number of parameters, less complexity, higher recognition accuracy for growth stages and nutritional status of tobacco plants, and is suitable for intelligent monitoring of tobacco plant growth in fields.