Raw seismic data contains noise from instrument errors, ambient vibrations, and multiples (reflections bouncing multiple times). Processing steps include denoising, gain correction, velocity analysis, normal moveout correction, and stacking. These operations enhance reflections from target interfaces while suppressing noise, producing final seismic images ready for interpretation.
From reflection seismic survey design, you understand how sources and receivers are arranged to record waves bouncing off subsurface interfaces. But what comes out of the field is not a clean image — it is a massive collection of wiggly traces full of noise, artifacts, and geometric distortions. Seismic data processing is the sequence of operations that transforms this raw data into an interpretable cross-section of the subsurface. Think of it as developing a photograph from a film negative: the information is in there, but it takes careful processing to reveal it.
The first steps address basic data quality. Gain correction compensates for the fact that seismic waves lose energy as they travel — deeper reflections arrive with much smaller amplitudes than shallow ones, so the traces are scaled to make reflections at all depths visible. Frequency filtering removes noise outside the useful signal band: low-frequency ground roll (surface waves generated by the source) and high-frequency random noise are attenuated using bandpass filters. Bad traces from malfunctioning receivers are identified and removed (a process called editing or trace killing).
The central processing step is normal moveout (NMO) correction and stacking. In a common midpoint (CMP) gather — all traces that share the same reflection point — the same reflection arrives at different times depending on the source-receiver offset. For a flat reflector, the travel-time curve is a hyperbola: traces at larger offsets record the reflection later because the wave travels a longer path. Velocity analysis determines the seismic velocity that best flattens this hyperbola. Once the correct velocity is found, NMO correction removes the offset-dependent time delay, aligning the reflection horizontally across all offsets. The corrected traces are then stacked — averaged together — which dramatically improves the signal-to-noise ratio because coherent reflections add constructively while random noise cancels out. A stack of 50 traces improves the signal-to-noise ratio by roughly a factor of 7.
After stacking, additional steps address remaining artifacts. Multiple suppression removes reflections that have bounced more than once between interfaces (such as the sea floor in marine data) — these multiples masquerade as deeper reflections and must be identified and removed. Techniques include predictive deconvolution, which uses the repetitive nature of multiples to predict and subtract them, and Radon transforms, which separate multiples from primaries based on their different moveout velocities. The final processed section — a stacked, filtered, deconvolved image — shows the subsurface as a series of reflection events positioned at the correct two-way travel time. Converting this to true depth and correctly positioning dipping reflectors requires migration, which is covered in the next topic in this sequence.