Abstract: Spatial predictions of seabed sediments based on samples in the Australian Marine Samples (MARS) database provide environmental baseline information used in the management of Australia’s marine jurisdiction, offshore resource development, and marine protected areas. The MARS database holds 14204 samples, distributed unevenly across the Australian Exclusive Economic Zone (AEEZ). As such, interpolations of seabed samples are often required for understanding target regions between collection sites. Data quality in the MARS database also varies, with data quality control resulting in the exclusion of over 7,000 samples, most of which were dredged samples. Dredged samples are thought to lower the accuracy of the spatial predictions produced from the database. In this study we examined whether these excluded dredged samples should for spatial predictions by assessing whether the dredged samples decreased or increased the accuracy of the resulting spatial predictions of seabed mud content, and whether the dredging method used altered the accuracy of the resulting prediction.

We confined our analyses to two contrasting areas in the AEEZ: the Southwest Region (407 total samples; 150 dredged) and Petrel Region (534 total samples; 297 dredged). We compared the accuracy of interpolated surfaces of mud content generated from QC’d samples with surfaces generated with samples from different dredge types (benthic, pipe, chain bag and unspecified). In the Southwest region samples included 73 benthic, 19 pipe, 41 chain bag, and 17 unspecified dredged samples. In the Petrel region samples included 46 pipe and 251 unspecified dredged samples. Spatial predictions of seabed mud content were made using Inverse Distance Weighting (IDW) and Ordinary Kriging (OK). Predictive errors were assessed based on leave-one-out cross-validation in terms of relative mean absolute error (RMAE). The effects of dredged samples on the predictive error were analysed using paired Mann-Whitney tests.

For sediment samples in the Southwest region, including benthic dredge samples in the prediction reduced the accuracy of IDW by 1.79% in terms of RMAE. Including chain bag dredge samples increased the accuracy of IDW with a small difference in RMAE of 0.47%. The difference in accuracy of IDW was insignificant for the unspecified type samples, all dredged sample types and pipe dredged samples. Including samples with all-dredged type and unspecified type dredge samples improved the accuracy of the OK predictions over the QCed samples in the Southwest region with a small effect on RMAE of 0.68% for all-dredged samples and 0.65% for unspecified type dredged samples. The difference in accuracy of OK was insignificant for benthic dredged, pipe dredged and chain bag dredged samples in the Southwest region. No significant effects on the accuracy of IDW in the Petrel region were found for all the tested dredged sample types (unspecified type, pipe dredged samples, and all-dredged samples). The difference in accuracy of OK in the Petrel Region for samples with all dredged and with unspecified dredged samples was insignificant. Including pipe dredged samples increased the accuracy of OK in the Petrel with a negligible effect on RMAE of 0.02%.

In summary, the inclusion of dredged samples produced minimal effects on the accuracy of spatial predictive models. Effects were not consistent across region or dredge type, but findings showed dredged samples had surprisingly little effect on the accuracy of the predictions Including the dredged samples would require reprocessing the spatial predictions for the AEEZ based on the MARS database, so as the findings are only based mud sediment data from two regions caution should be taken to generalise these findings to other sediment types and to the entire Australian marine margin. Further testing is required to verify these findings for other regions and sediment types and identify whether future spatial predictions can include some of this extra data without a loss in accuracy.

This paper was submitted/presented at the 22nd International Congress on Modelling and Simulation (MODSIM2017) conference, 3-8 December 2017