大多数偏移方法是以波动方程的各种单向近似法为基础,一个明显的例外是有限差分逆时深度偏移的算法。与其它的偏移算法相比,由于这种方法需要庞大的计算机设备,所以其最初应用只局限于二维资料的合成处理。因此,对于用波前相长干涉进行实际资料叠后偏移和复杂三维构造成像来说,此方法的潜力一直未能发挥出来。为了避免网格位移和数值的不稳定性,有限差分深度偏移受到与正演模拟方法所遇到的相同条件的支配。对于野外资料,这种方法可能需要在空间上和时间上都进行内插。因此,此方法一个能够发挥其潜力的事实是:在正演模拟中,尽力产生准确的反射信号;而在偏移中,主要目的是根据预先记录的信号完成成像,这些信号可能是在不太严格的条件下获得的。的确,我们的研究表明,沿水平方向(在此方向减少或排除使用内插道)时,对网格位移现象,即使没有硬性规定,也可以准确地成像。在三维资料的情况下,消除内插道的步骤使得计算工作量减少了50%或更多,存储的需求量也相应地减少了。沿垂直方向,使用不均匀网格,能够获得更高的效率,而在此方向,随输入资料速度和频率成分的变化,适合于向下传播子波的延展与压缩。这些步骤使进行大的三维数据体高分辨率偏移需要的时间减少了几小时或更多。减少的时间取决于使用的机器及输入数据的多少。本文给出了在大型并行计算机上进行大规模勘探三维野外数据处理的两个应用实例,把我们的处理结果与使用 Hale-McClellan算法得到的结果进行了比较。 Most offset methods are based on a variety of unidirectional approximations of wave equations, with one notable exception being the algorithm for finite-difference time-dependent depth migration. Compared with other offset algorithms, the initial application of this method is limited to the synthesis of two-dimensional data because it requires huge computer equipment. Therefore, the potential of this method has not been demonstrated for actual data of post-stack migration and complex three-dimensional imaging using wavefront constructive interference. Finite difference depth migration is governed by the same conditions encountered with forward modeling methods in order to avoid grid displacements and numerical instabilities. For field data, this method may need to be spatially and temporally interpolated. Therefore, one of the ways in which this method can exert its potential is to try to produce an accurate reflection signal in forward modeling, whereas in offset, the main purpose is to complete imaging from prerecorded signals, which may not be Obtained under strict conditions. Indeed, our research shows that grid displacement can be accurately imaged even in the absence of a rigid rule when it is oriented horizontally (reducing or excluding the use of inlets in this direction). In the case of three-dimensional material, the elimination of interpolation steps reduces the computational effort by 50% or more and the amount of stored demand is correspondingly reduced. Using a non-uniform grid in the vertical direction, higher efficiencies can be achieved, and in this direction, as the input data velocity and frequency components change, it is suitable for spreading and compressing the downward propagating wavelet. These steps reduce the time required to perform high-resolution large 3D volume shifts by a few hours or more. The amount of time to be reduced depends on the machine used and the amount of data entered. This paper presents two examples of large-scale exploration of three-dimensional field data processing on large-scale parallel computers, and compares our results with the results obtained using Hale-McClellan algorithm.