Deterministic analysis of coda waves has become one of the effective approaches to estimate inhomogeneous structure in the crust and uppermost mantle. Seismic array observation gives coherent wave trains, which are suitable for estimating scattering properties beneath the array. Local or regional seismographic network offers distribution of coda wave energy in space and time, which has an advantage of estimating scattering properties in a wide area below the station network. In this paper, several studies of the latter analysis are introduced. One of those analyses, proposed by Nishigami (1991), is to invert coda envelopes observed by a local seismographic network into spatial distribution of scattering coefficients. An observational equation is derived under the assumption of S to S single isotropic scattering and spherical source radiation, where observed data are the deviations of coda envelope from a master curve decaying smoothly. The unknown parameters are the relative scattering coefficients assigned to small blocks that divide the three-dimensional space. The observational equation can be solved by a recursive stochastic inversion. This analysis method applied to the central California data delineates deep structure of the San Andreas fault system as a structure with strong scattering. Several regions in central Japan were also analyzed by this method and detailed distribution of S wave reflectors was estimated below active volcanoes, Mt. Ontake and Mt. Nikko-Shirane. Weak scattering correlated with a seismic gap area was also recognized in the Hokuriku region. Revenaugh (1995a) proposed another method, callled Kirchhoff coda migration, which stacks the energy of single-scattered waves within the coda of teleseismic P waves observed by a local seismographic network. This method was applied to southern California and the following images were obtained, strong P to P scattering from the slab subducting beneath the Transverse Ranges at depths of 50-200km; P to Rg scattering correlated with topographic roughness; and P to S scattering from the aftershock area of the 1992 Landers earthquake. These two methods analyzing the seismic network data seem to be effective to estimate the heterogeneous structure in the crust and uppermost mantle. From the viewpoint of imaging the deep structure of active faults, the inversion analysis of coda envelopes from local earthquakes may be more effective. The variance reduction of the observed data, however, is low, for example, about 14% for the central California analysis. In the future, the simple models of scattering and source radiation assumed in the inversion analysis should be improved into more realistic ones to increase the model fitting. Forward modeling such as proposed by Obara (1997) will also be important to understand the whole shape of coda envelopes related to inhomogeneous distribution of scattering and attenuation intensity.