With the expansion of a city, the urban green space is occupied and the urban heat island effect is serious. Greening the roof surfaces of urban buildings is an effective way to increase the area of urban green space and improve the urban ecological environment. To provide effective data support for urban green space planning, this paper used high-resolution images to (1) obtain accurate building spots on the map of the study area through deep learning assisted manual correction; and (2) establish an evaluation index system of roof greening including the characteristics of the roof itself, the natural environment and the human society environment. The weight values of attributes not related to the roof itself were calculated by Analytic Hierarchy Process (AHP). The suitable green roof locations were evaluated by spatial join, weighted superposition and other spatial analysis methods. Taking the areas within the Chengdu city's third ring road as the study area, the results show that an accurate building pattern obtained by deep learning greatly improves the efficiency of the experiment. The roof surfaces unsuitable for greening can be effectively classified by the method of feature extraction, with an accuracy of 86.58%. The roofs suitable for greening account for 48.08%, among which, the high-suitability roofs, medium-suitability roofs and low-suitability roofs represent 45.32%, 38.95% and 15.73%. The high-suitability green buildings are mainly distributed in the first ring district and the western area outside the first ring district in Chengdu. This paper is useful for solving the current problem of the more saturated high-density urban area and allowing the expansion of the urban ecological environment.