The invention, which relates to the technical field of data mining, provides an image classification prediction method based on multi-task learning in sparse subspace. The method comprises: step one, constructing feature data according to a heterogeneous vision feature extracted from a training image set; step two, on the basis of correlation of tasks as well as a constitutive property of the heterogeneous vision feature, establishing an objective function of a multi-task classification algorithm of sparse subspace according to a training data set of the feature data; step three, carrying out solving by using an iterative optimization algorithm based on the objective function of the multi-task classification algorithm of the sparse subspace, thereby generating a multi-task classification model of sparse subspace; and step four, carrying out classification prediction on an image testing image set by using the multi-task classification model of sparse subspace. On the basis of combination of subspace correlation with a sparse mechanism, a multi-task learning mechanism is established; and the generalization performance of the classifier is improved and the generalization performance of image classification is enhanced.