Multibeam echosounders (MBES) are currently the best way to determine the bathymetry of large regions of the seabed with high accuracy. They are becoming the standard instrument for hydrographic surveying and are also used in geological studies, mineral exploration and scientific investigation of the earth's crustal deformations and life cycle. The significantly increased data density provided by an MBES has significant advantages in accurately delineating the morphology of the seabed, but comes with the attendant disadvantage of having to handle and process a much greater volume of data. Current data processing approaches typically involve (computer aided) human inspection of all data, with time‐consuming and subjective assessment of all data points. As data rates increase with each new generation of instrument and required turn‐around times decrease, manual approaches become unwieldy and automatic methods of processing essential. We propose a new method for automatically processing MBES data that attempts to address concerns of efficiency, objectivity, robustness and accuracy. The method attributes each sounding with an estimate of vertical and horizontal error, and then uses a model of information propagation to transfer information about the depth from each sounding to its local neighborhood. Embedded in the survey area are estimation nodes that aim to determine the true depth at an absolutely defined location, along with its associated uncertainty. As soon as soundings are made available, the nodes independently assimilate propagated information to form depth hypotheses which are then tracked and updated on‐line as more data is gathered. Consequently, we can extract at any time a “current‐best” estimate for all nodes, plus co‐located uncertainties and other metrics. The method can assimilate data from multiple surveys, multiple instruments or repeated passes of the same instrument in real‐time as data is being gathered. The data assimilation scheme is sufficiently robust to deal with typical survey echosounder errors. Robustness is improved by pre‐conditioning the data, and allowing the depth model to be incrementally defined. A model monitoring scheme ensures that inconsistent data are maintained as separate but internally consistent depth hypotheses. A disambiguation of these competing hypotheses is only carried out when required by the user. The algorithm has a low memory footprint, runs faster than data can currently be gathered, and is suitable for real‐time use. We call this algorithm CUBE (Combined Uncertainty and Bathymetry Estimator). We illustrate CUBE on two data sets gathered in shallow water with different instruments and for different purposes. We show that the algorithm is robust to even gross failure modes, and reliably processes the vast majority of the data. In both cases, we confirm that the estimates made by CUBE are statistically similar to those generated by hand.
