The invention discloses a multispectral remote sensing image crop straw off-field extraction method and system based on machine learning, and solves the problems that personal errors and measurement errors exist in the off-field area and progress of an existing straw off-field supervision mode, and accurate management and control cannot be realized on non-off-field area positioning. Comprising the following steps: S1, acquiring multispectral image data of a crop growth period, sketching and establishing a polygon sample set according to crop spectral information, establishing a crop classification model, and predicting the established crop classification model to obtain a crop distribution result; s2, acquiring multispectral image data after a crop harvest period, sketching polygonal samples of corresponding categories, and performing sampling treatment to obtain a polygonal sample set of an out-of-field plot; and S3, dividing the polygonal sample set of the out-of-field plot by utilizing five-fold division, training the obtained five groups of training sets-verification sets to obtain five base models, predicting the image data by using the base models, and determining the out-of-field plot through probability mean value fusion of prediction results.