To address the problem that the existing water-level detection method based on the gray image segmentation is susceptible to complex illumination conditions such as water surface flaring and reflection, and the problem of large measurement errors caused by the entanglement of floating objects during high flood periods, a deep-learning-based intelligent water-level monitoring method for staff gauge is designed. This method uses the three-class sample images of staff gauge,water surface, and floating objects, which are collected under different conditions and accurately labeled manually, to construct the data set, and then trains the deep fully convolutional neural network to perform the pixel-by-pixel classification prediction of staff gauge images. Finally, the pixel position of water line is detected in the semantic segmentation image, and transformed into the actual water level value. The test results show that the method can overcome the shortcoming of traditional methods in the image feature extraction, improve the adaptability of image segmentation to complex changing environments in the field, realize the recognition of measurement effectiveness, and achieve the purpose of intelligent monitoring of water level with the comprehensive uncertainty of measurement less than 3 cm.