Soy-RepViT: a lightweight deep learning framework for fine-grained recognition of soybean growth habits

Soybean growth habit is generally classified into determinate, indeterminate, and semi-determinate types, and its accurate recognition is important for breeding selection, high-density cultivation, and agronomic management. However, the semi-determinate type is intermediate between determinate and indeterminate types and shows continuous transitional characteristics with the indeterminate type in stem elongation, upper internode length, and pod distribution, making fine-grained discrimination between these two types difficult. Practical deployment also requires lightweight, efficient models. This study developed a lightweight fine-grained recognition framework for soybean growth habit from standardized plant images. The framework used RepViT-M1 as an efficient feature extraction backbone and integrated a multi-scale efficient channel attention module (MS-ECA) to enhance discriminative channel features, a Fine-grained Disambiguation Adapter (FDA) to recalibrate non-determinate features using class prototypes and reduce indeterminate/semi-determinate confusion, and a hierarchical classification head to incorporate the agronomic hierarchy between determinate and non-determinate types. The dataset comprised 944 original soybean plant images and was expanded to 4,709 images through offline augmentation. The model was evaluated using five-fold cross-validation and an independent test set. The proposed model achieved 97.45% accuracy on the independent test set, improving the RepViT-M1 baseline by approximately 3 percentage points. Accuracy for the indeterminate type increased from 91.30% to 97.83%, indicating that the model effectively reduced confusion between indeterminate and semi-determinate types. The model had 4.94 M parameters, required 0.82 GFLOPs, and reached 108.75 FPS, demonstrating a favorable balance between accuracy and efficiency. The proposed framework improves fine-grained soybean growth habit recognition at low computational cost and demonstrates potential for standardized, high-throughput phenotypic identification through on-device Android inference.