In terms of the issues that the accuracy of ship lock water level sensor is easily affected by the water quality and the poor adaptability of the traditional image detection methods, a waterline detection method based on semantic segmentation and coarse-to-fine strategy is proposed, and the calibration model with subsection control points is established to calculate the water level. Considering the characteristics of long-range dependence of waterline, the strip atrous spatial pyramid pooling module is proposed. To address the problem of inaccurate segmentation boundary, the multi-path aggregation upsample module is proposed and online hard example mining is introduced to improve the segmentation accuracy of the model. The improved semantic segmentation model is used to perform coarse detection, separates water and non-water regions in the low-resolution image, which is compressed from the original image, and obtains the coarse detection result of the waterline according to the binary mask output from the model. Then, crops the neighborhood of the coarse detected waterline in the original image and performs fine detection to obtain the fine detection result of the waterline. Finally, establishes the calibration model with subsection control points to build the relationship between pixel coordinates and world coordinates, and calculates water level by fine detection result. The experiments are implemented on the constructed water level dataset. Experimental results show that the improved semantic segmentation model increases mIoU by 2. 58% and 1. 98% for coarse detection and fine detection, respectively. The average pixel error of the fine detection result is 1. 89 pixel, which is 52. 3% lower than the coarse detection result. With manual observation of the water gauge as the benchmark, when the working distance of the camera is 23 m, the uncertainty of the measurement result with a confidence level of 95% is 0. 026 m. The proposed method has well adaptability for a variety of outdoor environments such as sunny, cloudy, rainy and snowy days, which provides an available method for water level detection of ship lock.