The water depth inversion of shallow reefs plays an important role for marine safety, marine engineering and marine military. The IKONOS satellite remote sensing image and the chart water depth data are analyzed and processed in this paper and a new neural network model is established by transmission bands ratio. The ratio of blue, green, red and near-infrared bands of IKONOS is used to calculate the water depth. Using neural network model, the water depth are inversed directly by the remote sensing image data without regard to other environmental factors (e.g. sea sediments, marine organisms, etc.). The non-linear relationship between the multi-spectral IKONOS data and the measured depth data can be established within the proposed model and higher inversion precision can be obtained compared with traditional regression model.