Advances in topo-bathymetric lidar systems now allow significantly higher data density, albeit in shallower water, than was previously possible.  This in turn leads to opportunities for better object detection and recognition, as well as semi-automated statistical data processing.  Design trade-offs in some of these systems can lead to many false detections, however, lowering the observation signal-to-noise ratio sufficiently to challenge any processing methodology; for the data here, having three quarters of the observations be non-bathymetric is not uncommon.

Conventional grid-based depth estimation developed for acoustic data can be successfully applied to this data in that the “noise” can be segregated into multiple depth hypotheses at each estimation node, but current methods for hypothesis selection fail in the face of the noise volume.

This paper proposes, therefore, an alternative scheme for depth hypothesis selection, based on machine learning techniques, that can successfully reconstruct the depth despite the levels of noise observed.  A variant of the CHRT algorithm is first deployed to form a fine-to-coarse, data-adaptive, variable-resolution grid estimate of depth.  A vector-quantized Hidden Markov Model (supervised) classification scheme is then used to sub-select from all hypotheses at a node only those that appear to be plausible seafloor reconstructions.  This minimizes the reconstruction ambiguity, allowing conventional techniques to be successfully applied, while also admitting a previously unavailable “no reconstruction” state in areas where no depth hypothesis appears valid.  This avoids, for example, reconstructing surface noise in areas beyond extinction depth.

The algorithm is illustrated using Reigl VQ-880-G data flown by NOAA Remote Sensing Division in 2016, which clearly demonstrates the improvement over conventional techniques.  In addition to improved performance, lower user workload, and data fidelity, this hybrid scheme is directly compatible with current acoustically-derived algorithms, with obvious advantages for software, procedural, and training outlays that derive from a common architecture.