Fram Strait data processing

The data used for the generation of the new Fram Strait bathymetry was extremly heterogeneous in terms of format, quality and processing stage making an automatic/operational processing of the whole data set impracticle. The processing applied can be divided into the following consecutive steps:

1. multibeam data processing in the swath domain (ping-depth-range coordinate system) for each cruise data file,
2. combination of cruise data files and data processing in the spatial domain (geographic coordinate system),
3. gridding of the cleaned depth data,
4. post processing (grid enhancing, contouring, mapping).

Under topic one the first step was the correction of obvious relative navigation errors. Typical error patterns include tracks slightly sliding out of course while GPS P-Dop (position dilution of precision) values are low and an abrupt jump back to the real track, e.g. after changes in the satellite constellation. In combination the motion sensor data (heave, pitch, roll) were checked for erroneous values.

Next, the crucial swath editing was performed. Automatic spike detection utilized local min./max. depth ranges and across/along track beam to beam slopes. For the automatic rejection of grouped erroneous depth values rejection was performed using multiples of average depth filters, allowed port and starboard across track distance ranges and acceptable port and starboard angles from nadir.

For areal quality control after every cleaning step a colour-coded shaded relief image of the track has been generated using a fast gridding algorithm. In erroneous areas insensitive to automatic detection manual swath editing was applied. The shaded reliefs turned out to be a valuable tool for identifying such areas. When refraction artifacts occured during swath editing adequate corrections were applied.

After the swath editing the data was transformed to the spatial domain (topic 2), enabling a joined analysis and processing of data from different cruises. First an automatical surface cleaning was performed to correct for remaining errors. The whole data set did undergo an iterative spatial tiling taking into account the desired tile size and number of included depth values. Following for each tile a 3rd order polynomial surface was computed utilising iteratively re-weighted least squares. Test criterias for the rejection were adequate threshold levels above and below the surface for residual testing and control of the re-weighting model for the residuals as a function of standard deviation. In some areas absolute residuals were defined for data rejection. The process was controlled by descriptive statistics and residual histogram analysis. It has to be noted, that only tiles with an approximately gaussian data distribution can reasonably be processed with the described method.

The last data cleaning step comprised manual editing on subset basis. In order to identify persisting outliers the depth values in critical areas were visualised in a system allowing 3d-views on the data from different perspectives. All in all the cleaning process as a whole was not so much straight forward as described, rather an iterative cycle allowing a "jump-back" in the processing chain if necessary.

In the calculation of depths in metres all swath data was transformed to "soundings" assuming a mean sound velocity in water of 1500 m/sec, that is, to uncorrected depth. The Seabaem system directly output soundings. Hydrosweep, in contrast, gives full information on signal run times, signal strength, angles, and so on. No local sound velocity profiles have been applied to either the Seabeam or Hydrosweep data: in deep sea studies, experience has shown that it is easier to fit different data sets together accordingly when working with uncorrected soundings. One reason is, that system calibration errors often overbalance sound velocity differences. However, local sound velocity profiles, e.g. from CTD measurements, can be applied at later processing stages, if necessary even on the raster DTMs or parts of them.

The cleaned depth data had to be gridded in order to compute a DTM (topic 3). Standard gridding techniques, e.g. IDW, spline, kriging, don't account for the data characteristics of multibeam systems, particularly the decreasing data quality from the center to the outer beams. Hence a weighted gridding was applied taking into account the local beam footprint size of a specific sounding, the associated grazing angle and distance to the adjacent grid nodes to be interpolated. Sounding weights decreased with increasing distance to the grid node and decreasing grazing angle. Grids in resolutions of 100 m, 200 m and 500 m were computed. The grids have nodata values at places where no sounding footprint overlappes a grid node. This occured at several places within the study area, particularly when the survey lines were not planned with sufficient overlapping, or when outer beam rejection due to bad data quality during depth editing led to swath narrowing. In both cases narrow elongated nodata patterns with varying width and length parallel to the survey track lines form. Typical widths vary in the range of tens to some hundred metres. In order to fill the nodata patches a triangulated irregular network (TIN) has been established on the basis of the grid nodes containing depth values. Simple re-gridding of the TIN led to linearly filled nodata patches.

Surface shading is a sensitive quality control tool and visualized some systematical effects still remaining in the gridded data, particularly line-shaped features parallel to the survey lines. These are typical for grids interpolated from multibeam data sources, often developed in regions where outer beams overlap, especially in flat terrain. Even if absolute depth variation in such areas is almost neglectable these features have to be eliminated in order to extract "clean" DTMs and shades for morphology interpretation and contours suitable for cartographic output.

The post-processing (topic 4) comprised two steps. First the grid was processed with a distance weighted x-shaped focal average filter matrix to clear the "ghost lines". Following the complete grid was focally average filtered under consideration of the local terrain slope. To prevent steep areas being flattened from large filter matrices a slope classification was performed. The size and weighting scheme of the applied average filters was controlled by the slope class associated with each grid node. The table below shows the slope classes and respective filter matrices.