In order to realize automatic water level monitoring based on video images and solve the problems of poor environmental adaptability and low robustness of traditional video monitoring algorithms, a video water level monitoring method based on semantic segmentation is proposed. The improved DeepLabv3+ algorithm, combined with spatial attention mechanism, channel attention mechanism and edge refinement module, is used to segment the water level scale image to extract horizontal coordinate, and the actual water level value is calculated according to the linear interpolation of camera calibration results. The experimental results show that the average intersection ratio of the proposed algorithm on the water level scale dataset reaches 97.18%, which is better than DeepLabv3+ and BiSeNet (Bilateral Segmentation Network) semantic segmentation algorithms. The average pixel error rate of the proposed algorithm is 0.76%, and the error of water level reading is less than 1 cm in the measured environment. Compared with the existing traditional image processing water level monitoring algorithm and water level monitoring algorithm based on deep learning, the proposed algorithm has stronger environmental adaptability, higher robustness, more accurate reading, and can achieve more accurate automatic water level monitoring.