Table 2.
Different metrics on the contribution of variables to the best models
| Variable | Relative contribution | Permutation importance | Jackknife training gain with only variable | Jackknife training gain without this variable |
|---|---|---|---|---|
| Broad-scale modeling | ||||
| ABR | 75.10 | 45.45 | 1.05 | 1.05 |
| B10 | 9.01 | 28.17 | 0.64 | 1.25 |
| B12 | 3.86 | 7.34 | 0.59 | 1.34 |
| B17 | 4.63 | 6.45 | 0.41 | 1.34 |
| B15 | 2.34 | 2.28 | 0.09 | 1.34 |
| LAND | 3.70 | 1.57 | 0.20 | 1.33 |
| B4 | 1.33 | 8.71 | 0.11 | 1.34 |
| Fine-scale modeling | ||||
| DIS-ROCK | 34.36 | 42.86 | 0.97 | 1.71 |
| SLOPE | 31.68 | 14.36 | 1.16 | 1.53 |
| ELEV | 25.50 | 28.49 | 1.02 | 1.68 |
| DIS-URBAN | 8.44 | 14.27 | 0.30 | 1.73 |
All values are averages of the 50 replicates of the best models. The relative contribution is obtained from the increase of the regularized gain when each variable is added to the model. The permutation importance is obtained by randomly permuting the values of that variable among the training points and measuring the decrease in training AUC produced by the permutation. The Jackknife training gain with only variable is the training gain that the model achieves when using only that variable, and the Jackknife training gain without this variable is the training gain that the model achieves when using the rest of variables except that one. Consequently, in the first three metrics larger values indicate higher contribution of variables to the model, while in the last one, the lower values indicate greater importance of variables to the model.