As a supplementary means of acoustic measurement, satellite depth inversion plays an important role in disputed islands and reefs. However, due to the lack of measured water depth data and the limitation of the number of visible light segments in multi-spectral images, the traditional empirical model and semi-analytical model can not be used. Aiming at the remote sensing images with only three visible bands, an adaptive empirical semi-analytical model without measured water depth data is proposed. The new model is based on the partial pixel water depth obtained by the joint semi-analytical model and two empirical models, and the cooperative adaptive linear ratio model can determine the final result. The Sentinel-2 image (spatial resolution is 10 m) is used to test the AESM model, Log_ratio model and Log-ratio and Semianalytical (L-S) model in Ganquan island and Langhua reef, and the calculated results are compared with the water depth data. The results show that the root mean square error (RMSE) of the new model is 1.14 m in Ganquan island and 1.55 m in Langhua reef, respectively. The inversion accuracy of the new model is slightly better than that of the Log_ratio model trained with 200 water depth data. Compared with the L-S model, which also does not require water depth data, the RMSE of the new model is reduced by 0.12 m and 1.25 m, respectively.