Recent advancements in three-dimensional (3-D) facies modeling of outcrops have enabled their prompt reconstruction through drone photogrammetry. Such 3-D outcrop models facilitate the understanding of spatial lithofacies distributions and serve as valuable tools for quantitative analysis. Nevertheless, challenges persist in identifying facies within the digital models of large-scale outcrops. To address this issue, this study proposes a method for automatically recognizing outcrop facies, using a convolutional neural network (CNN) model applied to a 3-D point cloud of an outcrop. The process involves constructing a 3-D outcrop point cloud through drone photogrammetry and translating the point cloud into a series of two-dimensional images containing the color and roughness properties of the outcrop surface. The CNN model was trained using manually annotated data sets with five distinct classes: pebbly mudstone, sedimentary block, vegetation, beach, and topsoil. The trained model accurately estimated outcrop facies in unknown test images, with a probability as high as 86.4% in terms of -score. Furthermore, experimental results revealed that considering roughness information significantly improved lithofacies classification accuracy. Our method was applied to the outcrop of a mass-transport deposit exposed in the Upper Cretaceous to Paleocene Akkeshi Formation along the Esashito coast of Hokkaido Island, northern Japan. The trained CNN model effectively reconstructed the 3-D facies model, which substantially agreed with the actual spatial distribution in visual assessments. This novel approach offers a swift and precise means of reconstructing 3-D facies models even in large-scale or inaccessible outcrops, paving the way for quantitative analyses across diverse regions.