With some offshore waters in China as the study area, relying on the bidirectional long short-term memory (Bi-LSTM), we established a model for optimizing and inverting the Sentinel-2A image data on August 20, 2017, and the measured water depth data by a single-beam sounding system. Then, the water depth was compared with that of traditional inversion methods in terms of accuracy. The results show that the model proposed in this paper achieves the optimal evaluation indexes, with root-mean-square error (RMSE) and mean relative error (MRE) of 0.85 m and 18.93%, respectively. The MRE is 14%--29% higher than those of other methods. In particular, the RMSE in the 0--2 m offshore shallow water region displays a sharp rise of 0.68 m. Furthermore, 5-fold cross validation was used to test the network model cyclically, demonstrating the highest and most stable accuracy of Bi-LSTM. This model can invert coastal water depth with high accuracy and stability, and also apply to the depth inversion of regions without training samples, meeting the marine needs such as resource development, engineering construction, and environmental protection.