大豆粗脂肪含量检测中结合相似性测度的K均值蚁群优化波长选择方法

To overcome the limitation of traditional Euclidean distance, which ignores the correlation of spectral variables, this paper proposes a wavelength selection strategy that integrates K-means and ant colony optimization (ACO) with similarity measurements (simKACO). The cosine similarity or correlation coefficient can constrain the clustering structure to highlight the wavelength region highly correlated with the target variable. Simultaneously, this metric guides the ant colony algorithm to heuristically search the key feature clusters to achieve efficient screening of wavelengths related to fat content in soybean near-infrared spectroscopy. The results indicated that the cosine similarity is the most suitable similarity measurement for simKACO. The R² of the KcosACOcos-PLS model's test set can reach 0.8665, demonstrating superior predictive performance and generalization capability compared to PLS models established using other feature selection methods (UVE, IRF, GA, ACO). This research provides a reference method for detecting food quality.