Earthquake events from the Indonesian subduction zone recorded in northern Australia show a long and high-frequency coda associated with both P and S waves. Regional events recorded by Warramunga array in northern Australia can separate out wave propagation through the mantle by focusing on the coherent signal across the medium-aperture array. Most of the incoherent wave components result from structures in the vicinity of the array with small-scale lengths of 1-2 km or smaller. The coherent phases with relatively rapid changes in waveforms are associated with the scattering of seismic waves by crustal and mantle heterogeneity, but in some case can be related to structural effects near the source. As the depth of the source increases, the coherent portion of the seismic wavefield tends to become much simpler, which suggests that the heterogeneity tends to weaken at depth with larger-scale length. We compare the coherent signal features of earthquakes from the Indonesian subduction zone that have occurred in recent years with those in the early 1980s, first studied by Kennett （Phys Earth Planet Inter 47： 319-332, 1987）. The general characteristics of the coherent signal variation with depth in recent years are the same as those observed in 1980s, but the variations are larger. This change suggests a stringer var- iation in heterogeneity with depth than before, which may bear important information about the dynamic processes and evolution of the crust and upper mantle.
Numerical solver using a uniform grid is popular due to its simplicity and low computational cost, but would be unfeasible in the presence of tiny structures in large-scale media. It is necessary to use a nonuniform grid, where upsampling the wavefield from the coarse grid to the fine grid is essential for reducing artifacts. In this paper, we suggest a local refinement scheme using the Fourier interpolation, which is superior to traditional interpolation methods since it is theoretically exact if the input wavefield is band limited.Traditional interpolation methods would fail at high upsampling ratios(say 50); in contrast, our scheme still works well in the same situations, and the upsampling ratio can be any positive integer. A high upsampling ratio allows us to greatly reduce the computational burden and memory demand in the presence of tiny structures and large-scale models, especially for 3D cases.
