The metrics table reports sensitivity and specificity at the default 0.5 cutoff, which is rarely right when the classes are imbalanced — and they are here by construction, since a 90th-percentile abundance threshold makes only 10% of stations patches. At 0.5 a random forest will call almost nothing a patch, so sensitivity looks far worse than the model can actually achieve.
Details
This is the cutoff the original pipeline was reaching for with
metric.binary = 'ROC' when it binarised its projections.
References
Allouche O, Tsoar A, Kadmon R (2006). Assessing the accuracy of species distribution models: prevalence, kappa and the true skill statistic (TSS). Journal of Applied Ecology 43(6), 1223-1232. doi:10.1111/j.1365-2664.2006.01214.x