To realize intelligent water management and control and flood disaster early warning, it is necessary to accurately sense the change of water level information in real time. Because the prior technology cannot meet the requirements of water level identification in complex and harsh environments such as night, fog, rainy day, floating object occlusion, light shadows, etc. , an intelligent water level detection method based on improved YOLOv5 and RankSE was proposed. Firstly, the YOLOv5 algorithm was improved by the multi-level feature fusion method which strengthens small-scale features, to strengthen the ability of capturing small targets. Secondly, integrating the RankSE module further enhances the perception of small targets. Finally, a new solution of water level elevation was proposed, which can obtain accurate water level elevation information only by using part of water gauge anchor frame information, which greatly improved the robustness of the detection method. The research results show that the accuracy of water level detection in this paper reached 98. 5%, which was 8. 4% higher than the original algorithm. The water level elevation could be automatically and accurately identified in complex and harsh environments. The maximum error was only 0. 11 m. The research results effectively improve the accuracy of water level detection in complex and harsh environments.