As an important nutritional component in milk, fat is an important index to evaluate milk quality. Hyperspectral image technology can provide tens to thousands of bands of data and can reflect the subtle spectral differences of different components in milk. On the other hand, there is often a strong correlation between adjacent bands, which increases the amount of calculation and easily causes problems such as dimension disaster. Therefore, it is very important to select bands for hyperspectral data. This paper proposes a PLS-ACO feature band selection method combined with a genetic algorithm to form a new feature band selection method of PLS-ACO-GA. The two methods proposed in this paper are based on ant colony optimization. The absolute value of the regression coefficient of the PLS regression model is the main basis for evaluating the importance of wavelength, which is used as the heuristic information of ant colony optimization. Ant colony optimization is used for intelligent search, combined with genetic algorithm to produce more excellent characteristic band combinations. To avoid that pls-aco algorithm only obtains the optimal local solution. The optimal band combination can better reflect the information of fat composition in milk. By calculating the wavelength contribution rate, the optimal band combination is selected and compared with the spectral feature selection methods of genetic algorithm, cars algorithm and basic ant colony optimization. Finally, the prediction effects of the PLS regression model under different feature selection methods are compared. PLS-ACO, PLS-ACO-GA, CARS, GA and ACO screened 18, 16, 40, 43 and 42 characteristic bands in the spectrum of milk samples, respectively. The PLS prediction model after the PLS-GA-ACO screening band has the best effect. The prediction sets R²_P and RMSEP are 0.997 6 and 0.062 2 respectively, followed by PLS-ACO, and the prediction sets R²_P and RMSEP are 0.997 0 and 0.077 8 respectively. PLS-ACO and PLS-ACO-GA reduce the number of characteristic bands and improve the accuracy of the model. MLR, RFR and PLS regression prediction models are established based on PLS-ACO-GA data after characteristic band selection. The R²_P and RMSEP of the MLR prediction model are 0.997 6 and 0.062 3 respectively. R²_P and RMSEP of the RFR regression model were 0.999 9 and 0.003 0 respectively, and R²_P and RMSEP of the PLS regression model were 0.997 6 and 0.062 2 respectively. RFR model performs best among the three regression prediction models. The results show that hyperspectral technology can detect the fat content in milk, which provides a new, rapid and non-destructive method for the detection of fat content in milk.