Abstract: In view of the problems of large errors and lack of universality in water level detection, a new two-stage deep learning based intelligent water level detection method was proposed. The method improved the feature fusion network of the existing YOLOX-S algorithm by introducing ASFF module to improve the intensity of feature information fusion, and optimized the binary cross entropy loss (BCE Loss) by using a more flexible polynomial loss (Poly Loss), forming an improved YOLOX-S model. Combined with traditional image processing technology, a two-stage intelligent water level recognition method based on the improved YOLOX-S model for standard double-color water level and water level "E" scale recognition was established, which effectively improved the accuracy of water level detection. Experimental results show that the average recognition rate of the first stage water level and the second stage water level "E" scale reaches 98.94% and 99.86% respectively. Moreover, the average error in calculating the water level is less than 0.6 cm and the range error is less than 0.8 cm, reducing by 1.98 cm and 3.22 cm respectively compared to traditional typical water level detection methods. The proposed method realizes high-precision intelligent recognition of water level and provides effective technical support for flood control and drought relief decision-making.