Multi branch model based on cross scale feature fusion for wheat seedling variety recognition

Accurate identification of wheat varieties at the seedling stage is crucial for maintaining seed purity and optimizing field management. However, the subtle phenotypic variations among seedlings present a significant challenge for visual recognition. To address this, we propose SeedlingNet, a novel deep learning model specifically designed for fine-grained wheat seedling variety classification. The core innovations of SeedlingNet include: The Kolmogorov-Arnold-based Convolutional Attention (KCA) mechanism, which dynamically enhances feature representation by replacing static activation functions with learnable, adaptive ones; A multi-scale feature fusion architecture that integrates hierarchical features to capture both global and local characteristics. We established a comprehensive image dataset of 13,600 images representing 17 wheat varieties at the early growth stage. Experimental results demonstrate that SeedlingNet achieves a remarkable classification accuracy of 99.26%, outperforming traditional machine learning methods and mainstream deep learning models. Ablation studies confirm the significant impact of the KCA module and the multi-scale fusion structure on the model's performance. This research provides an effective, non-destructive tool for early-stage variety identification, with strong potential for precision agriculture applications.The dataset is licensed in Zhang, Wenbo (2025), "Seedings", Mendeley Data, V1, doi: 10.17632/f8ykx4sz6w. 1.