Vegetation restoration potential mapping (VRPM) is of great importance for regional ecosystem restoration planning after long‐term land degradation or short‐term disasters. However, there are some problems to be solved in the current models for evaluating the potential of vegetation restoration. First, the models for VRPM are mostly established based on a knowledge‐driven index system. Although this kind of model is logically rigorous, it relies too much on expert knowledge and is relatively inefficient, especially for large‐area vegetation restoration assessments. Second, because of the spatial heterogeneity, as well as the absence of important indicators, the traditional global‐based evaluation models are difficult to adapt to the entire study area. In this study, an improved data‐driven approach, that is, a sliding‐window based similar habitat potential model, is developed for VRPM. The advantages of the new model include: (a) it is more efficient in determining the importance of each influencing factor without resorting to expert knowledge; (b) it is more locally adaptive than the global model because it performs sample training, rule building, and vegetation restoration potential calculation in each current local window. We provide a case‐study to show the modeling process and result interpretation of the new model.