Monocular depth estimation is a critical task that infers scene depth information from a single image. It is widely applied in fields such as autonomous driving, medical imaging, and defense. Deep learning methods have significantly enhanced the representational capacity and prediction accuracy of these models, particularly excelling at handling complex scenes, multi-scale features, and dynamic objects, which are challenging for traditional methods. This paper systematically reviews monocular depth estimation methods based on deep learning, beginning with an introduction to the fundamental technical process of monocular depth estimation. Then, based on the type of supervision, deep learning methods for monocular depth estimation are categorized into three groups: supervised learning methods are reviewed in terms of network structure, auxiliary information, loss functions, and depth discretization; unsupervised methods are summarized based on cues such as image pairs, masking, visual odometry, auxiliary information, and generative adversarial networks; semi-supervised methods are explored with respect to image pairs, semantic information, and generative adversarial networks. Subsequently, the paper outlines the main monocular depth estimation datasets and commonly used evaluation metrics, and lists the quantitative evaluation results of several methods on these datasets. Finally, the paper discusses application examples of monocular depth estimation based on deep learning, and highlights the major challenges and potential development directions for the future.