Accurate perception of water level changes is one of the key segments to achieve precision water affairs control and flood disaster, but harsh scenes such as low illumination, haze, rain and snow, freezing, lighting, and waves bring a great challenge to water level accurate detection. Aiming at the problem of accurate detection of water level in existing methods, this paper constructs a Unet model fused with transformer residual channel attention mechanism (called “TRCAM-Unet”), then, a water lever intelligent detection method in harsh environments based on TRCAM-Unet is proposed. The key technologies include that: Multi-level feature fusion is achieved by full scale connection structure. The relevance of regional feature is strengthened by transformer module. Strengthening the extraction ability of useful information and weakening the interference of useless information by residual channel attention module. The experiments and practices of water level semantic segmentation in harsh scenes shows that TRCAM-Unet achieved 98. 84% MIOU scores and 99. 42% MPA scores, the maximum error of water level detection outside 150 meters was not above 0. 08 m, mean water level deviation (MLD) had only 1. 609×10 -2 meters, it is much better than the mainstream semantic segmentation models such as Deeplab, PSPNet, Unet. This study has important application value for water level accurate detection in harsh scenes and flood disaster early warning.