Substrate observations were assembled from the following sources: Canadian Hydrographic Service (CHS), Dive survey data, CHS marsh data, Natural Resources Canada (NRCan), and Remotely Operated Vehicle (ROV) surveys. These observations were mapped to common bottom type classes: rock, mixed, sand, or mud.

Using ArcGIS, the points were classified into training (2/3) or testing (1/3) data. The training subset is fully withheld from the model-building process and is used to evaluate each model’s performance. The random partitioning of the data into training and test subsets does not address the issue of spatial autocorrelation between observations.

The code to create the model follows the following steps:

A raster stack of environmental predictors is created. Point observations are overlayed onto the raster stack, and the values from the predictors are extracted to the points. Records with invalid BType4 values or with NA values from the predictors are removed. The training data are weighted according to their prevalence and are used to fit a random forest model using the ranger package (which supports case weighting). The fitted model is used to predict to the extent of the input environmental raster stack, classifying the entire area into rock, mixed, sand, or mud. The predicted surface is exported as a GeoTIFF raster file in the same resolution as the raster stack predictors (in this case, 100 metre resolution) and with the same projection. Evaluation statistics including Kappa, and accuracy by predicted class are generated using the withheld test data.

Manual steps after the model has been generated:

1) Reproject layer to EPSG: 3005 (R does not support writing the top-level EPSG code to the coordinate reference system information).

2) Create a raster attribute table with SUBSTRATE field for text classes.

3) Add a field for prevalence in the raster attribute table.

The bathymetry and its derivatives were created using the Arcpy module in python (see Scripts section in the metadata). The bathymetry layer was smoothed using a focal mean prior to generating the following derivatives: slope -> standard deviation of slope, curvature, and rugosity. This step was required as a method to reduce artefacts found in the derivatives. During development, these artefacts from the non-smoothed bathymetric derivatives carried through to the predicted substrate models. The non-smoothed bathymetry was used to generate the standardized BPI layers because this processing already involves processing with a neighbourhood of cells. As input to the random forest model, the original non-smoothed bathymetry was used. Ocean energy layers were also included – mean bottom ocean currents and mean tidal speed on the bottom. The SOG region uses the Salish Sea NEMO model (Allen) as source data for the ocean energy layers. The other 4 regions are sourced the the BC ROMS model (Masson). The original data for the Regional Ocean Modeling System (ROMS) has a 3 by 3 km grid resolution. These data were interpolated using Spline with Barriers (ESRI) and resampled to 20 m resolution rasters. See the scripts section for a link to the ROMS data processing. The source data for the NEMO model had higher resolution in the Strait of Georgia and so it was decided that this model would be used for the SOG region. They were first interpolated to a 40 m cell resolution (because of computational limitations) and then resampled to 20 m. The NEMO Tidal and Circulation layers were smoothed using Focal Statistics in ArcGIS with a 13 cell neighbourhood. Fetch was also included as a predictor. Fetch points were interpolated to the extent of the raster stack (specific methods and exact source data are unclear). DFO 20 m bathymetry layers (that included terrestrial elevation data) were used as the bathymetry source. The bathymetric position index (BPI) layers were created using the Benthic Terrain Modeler toolbox and were standardized after being calculated. It is important to note that when calculating BPI, the tool expects bathymetric data to have negative values associated with depths rather than positive.

Predictor Layers:

1: Bathymetry

2: Slope (bathymetric derivative) – degrees

3: Standard Deviation of Slope (bathymetric derivative)

4: Broad Bathymetric Position Index (bathymetric derivative) - Inner Radius: 25 - Outer Radius: 250

5: Medium Bathymetric Position Index (bathymetric derivative) - Inner Radius: 10 - Outer Radius: 100

6: Fine Bathymetric Position Index (bathymetric derivative) - Inner Radius: 3 - Outer Radius: 25

7: Curvature (bathymetric derivative; slope of slope)

8: Rugosity (bathymetric derivative; Arc-Chord rugosity)

9: Circulation

10: Tidal

11: Fetch