ABSTRACT

Satellite imagery with high spatial and temporal resolutions provides strong support for remote sensing bathymetry. Heterogeneous underwater environments such as complex bottom types represent the main factor affecting accuracy of satellite derived bathymetry. Segmenting imagery can enhance geospatial adaptability and improve the accuracy. Conventionally, satellite imagery can be subdivided based on bottom types confirmed by a priori knowledge gained from in situ observations (e.g. video and photos); however, this process poses challenges concerning some politically sensitive or unreachable areas. Besides, water depth in coral reef areas is generally shallow (depth range: 0 to 5 m) and complex, and it has proved difficult to improve bathymetric mapping accuracy with passive multispectral imagery in such areas using traditional models. With the spatial distribution of the residuals of discrete points to realize regional segmentation, this study developed an effective partition model for satellite derived bathymetry without the need for a priori knowledge of bottom types. WorldView-2 high resolution (2 m) imagery of Dongdao Island (China) was used in the study. The results of a practical application of the model revealed the following. (1) The developed partition bathymetry model reflected underwater topographical changes and features more clearly than a comparative global model, especially in certain areas of complex terrain. (2) The accuracy of the residual partition model (Mean Absolute Error (MAE): 0.96 m, Mean Relative Error (MRE): 11.04%, and Root Mean Square Error (RMSE): 1.36 m) is greater than the global model and the conventional partition model. (3) In different water depth intervals, the residual partition model has always the best accuracy. Subregion models corresponding to the proposed models are more stable and more accurate than those to the conventional partition model. In areas of shallow water (depth: 0 to 5 m), the MRE of a subregion-A model using the residual partition method is 14.09%, i.e. better than the global model (19.45%). The findings demonstrate that the developed model can largely overcome the problem of poor precision in shallow water associated with conventional models.