Intelligent water level recognition method integrating semantic segmentation with spatiotemporal information

To address the low accuracy and poor stability of water level recognition in complex natural environments, which result from illumination changes, alternation between wet and dry conditions, and background interference, an intelligent water level recognition method that integrated semantic segmentation with spatiotemporal information was proposed. An FCN-RCA model was constructed by combining residual connections, a convolutional attention mechanism, and an atrous convolution module; an image segmentation optimization method based on the flooding algorithm was introduced, and highly robust semantic segmentation and structural integrity optimization for the water body area were achieved. The segmentation results were mapped to the physical space of the slope domain via an affine transformation to build a virtual water gauge, enabling precise quantification of the water level. By leveraging the spatiotemporal continuity of video frames, temporal consistency optimization was applied to the water level boundary to enhance the stability and reliability of the recognition results. The case study at the Taleldesayi Station in Xinjiang demonstrates that the proposed method achieves a mean intersection over union (mIoU) exceeding 0.92 under various meteorological, hydrological, and illumination conditions, exhibiting excellent segmentation performance and accurate extraction of water level boundaries. The dynamic water level optimization method based on spatiotemporal information effectively ensures the stability and continuity of water level boundary recognition. Compared with the pre-optimization result, the absolute mean absolute error (MAE) and root mean square error (RMSE) reduce by 0.006 m and 0.004 m, respectively, while the Nash-Sutcliffe efficiency (NSE) increases by 0.024. In validation with long-term measured data, the MAE for water level recognition remains within 0.04 m, and the NSE reaches up to 0.997. These results confirm that the method can achieve centimeter-level water level recognition in complex environments, with good temporal consistency and engineering applicability.