The invention relates to a near-infrared spectrum quantitative modeling method based on integrated learning. The near-infrared spectrum quantitative modeling method comprises: determining an initial data set, and pre-processing data, wherein the initial data set contains a correction set for modeling and a prediction set for validation; dividing correction set samples into p classes according to a clustering strategy, randomly selecting one sample from each class to form the validation set of a sub-model, and forming the correction set of the sub-model from the remaining part; training the correction set of the sub-model by using a selected quantitative modeling method, carrying out statistics on the prediction error information of the prediction set in the initial data set through model output, and obtaining a weight function using the variance of prediction error as a parameter; repeatedly performing the previous steps, constructing a plurality of the sub-models, and weighing various sub-models according to the variance of the prediction error to obtain a quantitative correction model having strong robustness; and completing quantitative modeling by combining the whole quantitative correction model and the prediction set sample. According to the present invention, with the method, the output result of the whole model is closer to the true value.