Single-seed soybean variety identification using multi-point near-infrared spectral data based on multiple instance learning, gated attention mechanism and additive angular margin loss

Addressing challenges in soybean variety identification-such as susceptibility to interference in single-point near-infrared spectroscopy measurements, dilution of key differences in multi-point averaging, and weak inter-class distinctions at fine granularity-this paper proposes the multiple instance learning-multilayer perceptron-gated attention mechanism-additive angular margin loss network (MMGA-Net), an end-to-end classification model based on multi-point spectral data from individual seeds. Extending a multilayer perceptron architecture with the principles of multiple instance learning, the model represents the four point spectra collected from an individual seed as a bag-level sample and performs supervision at the bag level, thereby enabling holistic characterization of each seed. A gated attention mechanism is introduced to adaptively infer the contribution of each sampling point, effectively suppressing local anomalous interference while enhancing the robustness of the learned spectral representation. Meanwhile, an additive angular margin loss is introduced to enhance intra-class compactness and inter-class separability in the normalized feature space, thereby improving the clarity of the decision boundary. Across classification tasks involving 14 soybean varieties, the proposed model achieved accuracy, precision, recall and F1-score all exceeding 0.99, substantially outperforming conventional chemometric approaches and standard deep-learning models. This research provides robust technical support for stable multi-point spectral fusion and precise variety identification in single-seed analysis.