Abstract: Addressing problems of insufficient feature extraction and low recognition accuracy caused by small target size and similar color characteristics among diseases in sugarcane leaf diseases, an improved EfficientNetV2-S model has been proposed．Firstly, 5×5 convolutional kernel has been used to replace 3×3 convolution kernels from layers 12 to 17 in network, enlarging receptive field of convolution kernels to improve feature extraction abilities．Secondly, a global attention mechanism（GAM）has been introduced between input end and backbone network to improve model's ability to pay attention to diseased areas, further improving feature extraction ability．Finally, confidence label smoothing（CLS）algorithm has been combined with a loss function to optimize model．Experimental results showed that average accuracy of improved model reached 99.61% on test set．Compared with basic model, Top-1 accuracy, average precision, and F1-score have increased by 2.04, 1.97, and 0.87 percentage points respectively．Improved EfficientNetV2-S model performed well in task of identifying sugarcane leaf diseases, providing effective technical support for sugarcane leaf disease identification．