Enhanced Winter Wheat Seedling Classification and Identification Using the SETFL-ConvNeXt Model: Addressing Overfitting and Optimizing Training Strategies

Abstract

The growth status of winter wheat seedlings during the greening period is called the seedling situation. Timely and accurate determinations of the seedling situation type are important for subsequent field management measures and yield estimation. To solve the problems of low-efficiency artificial classification, subjective doping, inaccurate classification, and overfitting in transfer learning in classifying the seedling condition of winter wheat seedlings during the greening period, we propose an improved ConvNeXt winter wheat seedling status classification and identification network based on the pre-training–fine-tuning model addressing over-fitting in transfer learning. Based on ConvNeXt, a SETFL-ConvNeXt network (Squeeze and Excitation attention-tanh ConvNeXt using focal loss), a winter wheat seedling identification and grading network was designed by adding an improved SET attention module (Squeeze and Excitation attention-tanh) and replacing the Focal Loss function. The accuracy of the SETFL-ConvNeXt reached 96.68%. Compared with the classic ConvNeXt model, the accuracy of the Strong class, First class, and Third class increased by 1.188%, 2.199%, and 0.132%, respectively. With the model, we also compared the effects of different optimization strategies, five pre-training-fine-tuning models, and the degree of change in the pre-trained model. The accuracy of the fine-tuning models trained in the remaining layers increased by 0.19–6.19% using the last three frozen blocks, and the accuracy of the pre-trained model increased by 3.1–8.56% with the least degree of change method compared with the other methods. The SETFL-ConvNeXt network proposed in this study has high accuracy and can effectively address overfitting, providing theoretical and technical support for classifying winter wheat seedlings during the greening period. It also provides solutions and ideas for researchers who encounter overfitting.