The rupture characteristics of the 1983 Nihonkai-chubu (Japan Sea) earthquake (MJMA 7.7) were investigated using strong motion accelerograms recorded at 10 stations with epicentral distances of between 80 and 280 km. The main shock emitted a large amount of high-frequency seismic energy in two stages, forming two high-amplitude envelopes on the accelerograms. The time difference between S-wave arrival-times of the two events becomes larger as one moves clockwise in azimuth from north to south. Using the time differences, the second event was located 44 km NNE of the first one with the time of origin 26 s after the first event. The azimuthal variation in the amplitude ratio of the two events is consistent with the relative locations. The strong motion accelerograms, combined with the aftershock distribution and the source process time of 63 s obtained from long-period surface waves, suggest the following rupture model. The first event initiated at the southern end of the aftershock area and extended in a direction of N15°E. It stopped near the zone of low after shock activity west of Kyuroku Island. After a pause of about 10 s, the second event started at a place just north of the zone of low aftershock activity and extended in the same direction. A third event initiated near the place where the strike of the aftershock distribution changes from N15°E to N15°W. A low rupture velocity and the paucity of small-scale barriers are possible reasons for the low radiation of high-frequency energy from the third event. The above rupture characteristics appear to be closely correlated with the heterogeneous crustal structure as revealed by other geophysical and geological data in the source region. The observed S-wave accelerations of the first and the second events were jointly inverted for the source acceleration spectra, the attenuation coefficient of Qβ, and the amplification factors of the recording sites. The source spectra of the two events are almost the same, showing a rapid decay of amplitude at frequencies higher than 4 Hz. The Qβ increases in proportion to the frequency from 66 at 1 Hz to 1, 026 at 16 Hz. The local stress drops of subsources in terms of the stochastic source models are estimated to be 380 bars for the first and 340 bars for the second event. The estimates are model-dependent and can vary by as much as a factor of 1.5. The second event being more efficient than the first in radiating high-frequency seismic waves can be interpreted as having smaller subsources.