Plotkin, Rao, and Smith (SODA'97) showed that any graph with $m$ edges and $n$ vertices that excludes $K_h$ as a depth $O(\ell\log n)$-minor has a separator of size $O(n/\ell + \ell h^2\log n)$ and that such a separator can be found in $O(mn/\ell)$ time. A time bound of $O(m + n^{2+ε}/\ell)$ for any constant $ε> 0$ was later given (W., FOCS'11) which is an improvement for non-sparse graphs. We give three new algorithms. The first has the same separator size and running time $O(\mbox{poly}(h)\ell m^{1+ε})$. This is a significant improvement for small $h$ and $\ell$. If $\ell = Ω(n^{ε'})$ for an arbitrarily small chosen constant $ε' > 0$, we get a time bound of $O(\mbox{poly}(h)\ell n^{1+ε})$. The second algorithm achieves the same separator size (with a slightly larger polynomial dependency on $h$) and running time $O(\mbox{poly}(h)(\sqrt\ell n^{1+ε} + n^{2+ε}/\ell^{3/2}))$ when $\ell = Ω(n^{ε'})$. Our third algorithm has running time $O(\mbox{poly}(h)\sqrt\ell n^{1+ε})$ when $\ell = Ω(n^{ε'})$. It finds a separator of size $O(n/\ell) + \tilde O(\mbox{poly}(h)\ell\sqrt n)$ which is no worse than previous bounds when $h$ is fixed and $\ell = \tilde O(n^{1/4})$. A main tool in obtaining our results is a novel application of a decremental approximate distance oracle of Roditty and Zwick.