Water Level Detection Without Gauge Based on Squeeze-and-Excitation Module and Transformer

Real-time accurate water level data are indispensable in hydrology and water resource management. The accuracy of water level recognition is limited by factors such as weather changes, complex backgrounds, and water surface pollution; to address the above problems, firstly, the data set is preprocessed using Harmonization, which enriches the water surface and bank features in the data set and improves the quality of the data by making local data enhancement to the water surface or bank of the water level image, and using Harmonization to adjust the color and brightness characteristics of these regions to merge them with the background color; and then Construct the semantic segmentation model of the improved U-Net, and use the Transformer module to strengthen the correlation of the regional modules in multi-scale feature extraction to improve the recognition accuracy; meanwhile, combine with the SE Attention module to focus the feature extraction on the local key target region of the image, such as the intersection region of the water surface and the river bank, to enhance the expression ability of the improved network. The experimental results show that the segmentation accuracy of the improved segmentation model on a dataset of a water level site reaches 98.97% MIOU and 99.63% MPA, and the water level line pixel error (AEP) is as low as 1.18; the results of the study are of great application value in solving the challenges of accurate water level detection in various climate scenarios.