Fig 1.
German Bank study area with some of the input variables used in this study: the ground-truth data for the bottom types, the sea scallops observations, the bathymetry, the three backscatter derivatives and the six terrain attributes from Selection 1.
Table 1.
Selections of terrain attributes used to build the habitat maps and models.
The ID numbers refer to Lecours et al. [4] and allow finding the software and parameters with which the attributes were generated (see also S1 Appendix). Marker variables correspond to important variables; whether they were found on strong components (Sel. 1) or weak components (Sel. 4) is linked to the amount of topographic structure they accounted for. Variables with low cardinality (Sel. 2) did not have many different values, thus limiting their ability to explain slight variations in terrain morphology. Complex variables (Sel. 3) correspond to redundant variables. The terrain attributes identified by an asterisk were previously identified as potentially important [4]. The underlined attributes were recommended in [4].
Fig 2.
Map accuracies measured with (A) a kappa coefficient of agreement and (B) the overall accuracy.
Fig 3.
Comparison of the discrimination ability of the computed classifications with that of the classification computed using only bathymetry and the backscatter derivatives, based on the number of bottom types (maximum possible of 5) that were better discriminated.
Table 2.
Spatial similarity of the habitat maps and SDMs generated from Selections 2 to 7, compared to the map and model built from Selection 1.
A similarity of 90% indicates that 90% of the pixels were classified as the same habitat type in the two compared maps, or that 90% of the pixels were within ±5% of probability distribution in the two compared models.
Fig 4.
Performance and robustness of the 29 MaxEnt models.
Models in the top-left corner of the graph performed better and are more robust. Colour legend: Selection 1 (black), Selection 2 (blue), Selection 3 (red), Selection 4 (green), Selection 5 (purple), Selection 6 (orange), Selection 7 (white).
Fig 5.
Generalizability of the 29 MaxEnt models.
Models closer to the top-left corner are more generalizable as they performed well on the training data and replicated well to the validation data. See Fig 4 for colour legend.
Fig 6.
Percentage of variable contribution for the 29 MaxEnt models.
Only contributions greater than 5% are labeled.