基于分段注意力机制对大麦数据分类

Classification of Barley Data Based on a Segmented Attention Mechanism

摘要: 近红外光谱技术因其快速、无损等优势，在农产品品质分析中具有广泛应用，但高维、非线性光谱数据给多品种精细分类带来挑战。文章提出一种基于分段独立卷积注意力模块(SI-CBAM)的近红外光谱分类方法，用于埃塞俄比亚24个大麦品种的鉴别。针对大麦光谱在740~1070 nm范围内变化平缓、无明显吸收峰的特点，设计基于光谱导数分析的物理意义分段策略，将连续光谱划分为四个具有不同变化特征的子区间。每个子区间独立嵌入CBAM模块，通过通道注意力和空间注意力自适应强化局部判别特征，提升模型对细微光谱差异的感知能力。实验结果表明，SI-CBAM在测试集上分类准确率达0.8333，交叉验证准确率为0.8452，显著优于随机森林(0.3194)、支持向量机(0.7194)及全局CBAM模型(0.8000)。研究表明，结合光谱物理分段与注意力机制的策略，能有效提升近红外光谱在复杂分类任务中的判别性能与模型可解释性。

Abstract: Near-infrared spectroscopy, owing to its rapid and non-destructive advantages, finds extensive application in agricultural product quality analysis. However, high-dimensional, non-linear spectral data poses challenges for precise classification across multiple varieties. This paper proposes a near-infrared spectral classification method based on the Segmented Independent Convolutional Attention Module (SI-CBAM) for the identification of 24 barley varieties in Ethiopia. Addressing barley spectra’s characteristic gradual variation without distinct absorption peaks across 740~1070 nm, a physically meaningful segmentation strategy based on spectral derivative analysis was devised. This divides the continuous spectrum into four sub-intervals exhibiting distinct variation patterns. Each sub-interval is independently embedded within the CBAM module, where adaptive reinforcement of local discriminative features through channel attention and spatial attention enhances the model’s sensitivity to subtle spectral differences. Experimental results demonstrate that SI-CBAM achieves a classification accuracy of 0.8333 on the test set and a cross-validation accuracy of 0.8452, significantly outperforming Random Forest (0.3194), Support Vector Machine (0.7194), and the global CBAM model (0.8000). This research demonstrates that integrating spectral physical segmentation with attention mechanisms effectively enhances the discriminative performance and model interpretability of near-infrared spectroscopy in complex classification tasks.