The invention discloses a training image building method based on a field outcrop three-dimensional model, and the method comprises the steps: firstly selecting a field outcrop block, obtaining fieldoutcrop model data based on an unmanned plane oblique photography technology, and building a three-dimensional quantitative space model of a field outcrop; establishing a reservoir geological knowledge base, and counting reservoir skeleton model parameters; setting outcrop model virtual well points, generating X, Y and Z virtual well point coordinate information data files, virtual well point lithology distribution data and horizon depth data of different virtual wells, and establishing a three-dimensional grid skeleton model of a field outcrop area by utilizing three-dimensional modeling software based on the generated data; and simulating a field outcrop three-dimensional lithologic model through sequential indication based on a two-point statistical method of a variation function; calculating plane lithology contour line distribution data of different layers by using an arithmetic weighted average method, and establishing a training image three-dimensional space distribution model in combination with a reservoir geological knowledge base, so that the problem that a more accurate training image is difficult to obtain at present is solved.