Ambient Noise Interferometry

It seems counterintuitive that the constant, random trembling of the Earth—the ambient noise we typically try to filter out—could be the very source of profound geological insight. Yet, the revolutionary method of ambient noise interferometry achieves just that. It provides a way to image the Earth's interior and monitor its changes over time without relying on unpredictable earthquakes or costly, invasive artificial sources like explosions. This approach addresses the long-standing challenge of obtaining a continuous, high-resolution view of our dynamic planet.

This article will guide you through this remarkable technique. In the first chapter, ​​"Principles and Mechanisms"​​, we will explore the fundamental concepts, from the mathematical magic of cross-correlation that extracts a clear signal from noise, to the retrieval of the vital Green's function, which acts as a virtual seismic experiment. Following this, the chapter on ​​"Applications and Interdisciplinary Connections"​​ will demonstrate the power of this method in action. We will see how it is used to create detailed maps of the Earth's crust, function as a planetary-scale stethoscope for monitoring volcanoes and faults, and even find applications in urban environments, turning the rumble of a city into a geotechnical tool.

Imagine you are in a completely dark, cavernous room. You can't see a thing, but you want to map its shape. What do you do? You could clap your hands and listen to the echoes. The time it takes for an echo to return tells you the distance to a wall. Now, imagine the room is not silent. Instead, it’s filled with a constant, hissing static—a random, directionless cacophony of sound. It seems like an impossible task to hear any echoes in that din. Yet, the remarkable core of ambient noise interferometry is that this random noise is not a hindrance; it is the source of our information. By listening to the Earth's ever-present, subtle hum, we can create a picture of its interior as if we were setting off our own controlled explosions. Let's journey into how this seemingly magical feat is accomplished.

The central tool in our kit is a mathematical operation called ​​cross-correlation​​. Let’s return to our noisy room, but now we have two microphones, A and B, placed some distance apart. We record the hiss at both locations. The cross-correlation function, let's call it $C_{AB}(\tau)$, is a measure of how similar the signal at microphone A is to the signal at microphone B, when the recording from B is shifted in time by an amount $\tau$.

If a random sound wave happens to pass microphone A and then, a little later, microphone B, both will record a similar wiggle. When we slide the recording from B backward in time and compare it to A, we will find a moment, a specific time lag $\tau$, where those two wiggles line up perfectly. At this particular $\tau$, the correlation value will be high. This time lag, of course, is precisely the travel time of the sound between A and B.

Now, a single, random sound wave won't give us much. But in a field of random noise, waves are constantly crisscrossing our microphones from every possible direction. When we average the cross-correlation over a very long time, something amazing happens. The contributions from all the unrelated, incoherent waves—sounds that hit A but not B, or hit them in the wrong order—average out to zero. What remains, what gets reinforced, are the consistent echoes of waves that traveled the path between A and B. We are left with a clear signal emerging from the noise, with a peak at a time lag $\tau$ equal to the travel time. Even more beautifully, we also get a peak at $-\tau$, corresponding to waves that traveled from B to A. The cross-correlation function is two-sided, giving us the travel time in both directions!

This "echo" we retrieve is no ordinary signal. Under the right conditions, it is a profound physical quantity known as the ​​Green's function​​. You can think of the Green's function, often denoted $G(\mathbf{x}_A, \mathbf{x}_B, t)$, as the most elementary response of a system. It is the precise vibration you would record at location $\mathbf{x}_A$ if you were to give the medium a single, infinitely sharp "poke" (a delta-function impulse) at location $\mathbf{x}_B$ at time $t=0$. It contains everything there is to know about how waves travel from B to A, including their travel time, their change in shape, and their loss of energy due to attenuation.

The astonishing result of seismic interferometry is that the cross-correlation of a diffuse noise field, after some processing, becomes the Green's function. Specifically, the time derivative of the cross-correlation function $C_{AB}(t)$ is proportional to the Green's function from B to A, $G(\mathbf{x}_A, \mathbf{x}_B, t)$, minus its time-reversed version, $G(\mathbf{x}_A, \mathbf{x}_B, -t)$. This means that by passively listening to noise, we have created a ​​virtual source​​. It's as if we placed a source at station B and recorded the result at station A, without ever generating a signal ourselves.

This magic rests on a deep symmetry of nature known as ​​reciprocity​​. In most simple media, the path from A to B is the same as the path from B to A; formally, $G(\mathbf{x}_A, \mathbf{x}_B, t) = G(\mathbf{x}_B, \mathbf{x}_A, t)$. This fundamental principle, derivable from the wave equation itself, ensures that the virtual source experiment works both ways and is what makes the connection between passive noise and active-source experiments so elegant.

This retrieval of the Green's function is not guaranteed. It works under a critical assumption: that the ambient noise field is ​​diffuse​​. A diffuse field is one where, on average, wave energy is flowing equally in all directions. This is a state of ​​equipartition​​, where energy is evenly distributed among all possible wave types and propagation directions. Think of it as the difference between standing in a room with a single loudspeaker in the corner versus being inside a sphere of infinitely many tiny speakers, all chattering randomly. Only in the second case is the sound field truly diffuse.

In seismology, the main sources of ambient noise are ocean waves interacting with the seafloor, creating a global "hum." If these sources were perfectly distributed around our seismic stations, the noise field would be diffuse. What does this mean for our cross-correlation? It means the energy traveling from A to B is, on average, the same as the energy traveling from B to A. Consequently, the causal part of our correlation (for positive time lags $\tau > 0$) and the acausal part (for negative time lags $\tau < 0$) will have symmetric, equal-strength amplitudes.

But what if the noise is not diffuse? Imagine most of the noise comes from a distant coastline, arriving from one dominant direction. This ​​anisotropic​​ source distribution breaks the symmetry. If the dominant waves travel from A to B, we will get a strong signal in the causal part of our correlation but a very weak one in the acausal part. The cross-correlation becomes lopsided. Far from being a problem, this asymmetry is a powerful diagnostic tool, telling us about the directionality of the Earth's primary noise sources.

The leap from theory to practice relies on two powerful statistical ideas: ​​stationarity​​ and ​​ergodicity​​. A random process is ​​wide-sense stationary (WSS)​​ if its basic statistical properties, like its mean and variance, do not change over time. The Earth's hum, averaged over days or months, is a good approximation of a stationary process. This means the underlying statistics of the noise are consistent.

More importantly, we assume the noise is ​​ergodic​​. Ergodicity is the crucial property that allows us to substitute a time average for an "ensemble average." An ensemble average would require us to observe the noise fields in an infinite number of parallel universes and average them—clearly impossible. An ergodic process is one where a single, sufficiently long recording is representative of the entire ensemble. It means that by recording for a long time, the system explores all its possible states, and the time average converges to the true statistical average. It is ergodicity that gives us the license to take our single, long seismic record and, through time-averaging, have it reveal the underlying, deterministic Green's function. For this to hold mathematically, the noise process must be sufficiently "mixed" in time, a condition related to its autocorrelation decaying quickly enough.

We've talked about cross-correlation as a "shift, multiply, and sum" operation in the time domain. This is intuitive, but computationally slow. Fortunately, a beautiful mathematical theorem, the ​​Wiener-Khinchin theorem​​, provides an elegant and fast alternative in the frequency domain.

First, we use the Fourier transform to decompose our time-domain signals, $x_A(t)$ and $x_B(t)$, into their frequency components, $X_A(\omega)$ and $X_B(\omega)$. Each of these is a complex number for each frequency $\omega$, containing both amplitude and phase information. The theorem states that the Fourier transform of the cross-correlation function is simply the product of the individual Fourier transforms, with one of them being complex conjugated. This product is called the ​​cross-power spectral density​​ or ​​cross-spectrum​​, $S_{AB}(\omega) = X_A(\omega) X_B^*(\omega)$.

To get our time-domain correlation $C_{AB}(t)$, we simply perform an inverse Fourier transform on the cross-spectrum. This frequency-domain approach is vastly more efficient and allows us to manipulate the signal in powerful ways before transforming back to the time domain.

Real-world seismic data is not the perfect, stationary, Gaussian noise of our idealized models. It is contaminated by earthquakes, local human activity, and instrument glitches. The ambient noise spectrum itself is not flat; it has huge peaks at certain frequencies (the microseisms). To get a clean Green's function, we must first "tame" this raw data. This involves a multi-step processing workflow.

A crucial first step is precise ​​timing correction​​. Over days and months, station clocks can drift. Even a millisecond of error can ruin the coherent averaging process, so signals are resampled to a common, high-precision clock.

Next, we address the wild swings in signal amplitude. Several clever normalization schemes are used for this, all of which involve a fascinating trade-off: they sacrifice true amplitude information in order to stabilize and clarify the travel-time (phase) information.

• ​​Temporal Normalization​​: This involves dividing the signal at each moment by its slowly-varying local amplitude. This "tamps down" loud, transient events like earthquakes, allowing the quiet, persistent background noise to dominate the correlation.
• ​​Spectral Whitening​​: This is done in the frequency domain. We divide the spectrum of the signal by its own amplitude, effectively setting the amplitude at all frequencies to one while preserving the phase. This prevents dominant, narrow-band noise from overwhelming the correlation and results in a sharper, more broadband Green's function.
• ​​One-bit Normalization​​: This is the most aggressive technique. The entire seismic record is reduced to just its sign: $+1$ if the ground is moving up, $-1$ if it's moving down. Amazingly, for many types of noise, the cross-correlation of these one-bit signals still robustly recovers the correct travel times, a testament to how deeply the phase information is encoded.

A more advanced technique, ​​deconvolution interferometry​​, goes a step further. Instead of just correlating station A and B, it divides the cross-spectrum $C_{AB}(\omega)$ by the auto-spectrum of station B, $C_{BB}(\omega)$. Under ideal conditions, this can more effectively cancel out the spectral "color" of the noise source, providing an even cleaner estimate of the Earth's response.

The final piece of our puzzle reveals a breathtaking unity in physics. While most ambient seismic noise comes from oceans and atmosphere, some of it is truly the Earth's own internal thermal noise—the result of the random jiggling of atoms, just like the thermal noise in an electronic resistor. In this regime, seismic interferometry becomes a direct manifestation of the ​​Fluctuation-Dissipation Theorem​​.

This profound theorem states that the way a system dissipates energy when perturbed (a property encapsulated by the Green's function) is directly related to the spectrum of its spontaneous fluctuations when at thermal equilibrium. In other words, the random hiss of a system contains the blueprint for its deterministic response. The theorem predicts that for a system in thermal equilibrium, the noise cross-spectrum is directly proportional to the imaginary part of the Green's function, scaled by temperature and frequency. This connects the microscopic world of statistical mechanics to the planetary-scale observations of seismology, showing that the same fundamental principles govern the hiss of an amplifier and the hum of a planet. It is a powerful reminder that in the seemingly random noise of the universe, the deepest structures of reality are waiting to be discovered.

In the last chapter, we uncovered a most remarkable piece of magic. We learned that by listening patiently to the seemingly random, incoherent trembling of the Earth—the "ambient noise"—and performing a simple mathematical trick of cross-correlation, we can conjure up the signal that would have been recorded if one of our seismometers had been a source. We can, in essence, make the Earth speak to us on command, without ever needing the violent shout of an earthquake or the artificial bang of an explosion. We have retrieved the Green's function, the fundamental response of the medium between any two points.

This is a delightful and surprising result. But a good physicist, or any curious person for that matter, will immediately ask: "So what?" What is this trick good for? What new things can we do now that we have this power? It turns out that this is not merely a clever curiosity; it is a key that has unlocked a vast array of new ways to see, probe, and understand our world. We have moved from simply developing a new kind of "microphone" to having a full-fledged imaging system, a planetary-scale stethoscope, and even a toolkit for civil engineering. Let us now embark on a journey through some of these fascinating applications.

The most direct application of our newfound ability is to create maps of the Earth's interior, a technique known as seismic tomography. The principle is simple, at least in concept. The Green's function we retrieve between two stations is essentially a recording of a wave traveling from one to the other. By measuring the travel time, we can deduce the speed at which the wave traveled. If we do this for many pairs of stations, crisscrossing a region with a dense web of paths, we can build up a map of the seismic velocities in that region, much like how a CT scan uses X-rays from many angles to build a 3D image of the human body.

But there's a subtlety. The surface waves that dominate the ambient noise field are dispersive, meaning waves of different frequencies travel at different speeds. A low-frequency wave, with its long wavelength, "feels" deeper into the Earth, while a high-frequency wave is sensitive only to the shallow structure. This is a wonderful gift! It means that by dissecting our retrieved signal into its constituent frequencies, we can map the Earth's velocity structure at different depths.

The technique for this is a beautiful piece of signal processing called Frequency-Time Analysis (FTAN). We take our cross-correlation signal and pass it through a series of very narrow "filters," each tuned to a specific frequency. This is like having a set of tuning forks that only resonate with one particular note. For each filtered signal, we look at when the energy packet arrives. This arrival time corresponds to the group velocity—the speed of the energy for that specific frequency. By doing this for a whole range of frequencies, we can measure a dispersion curve, which shows how velocity changes with frequency. A key insight here is the choice of filter; scientists often use Gaussian-shaped filters because they provide the best possible compromise between knowing the exact time and the exact frequency of the wave, a limit imposed by the fundamental uncertainty principle of physics.

Once we have a collection of these dispersion curves for many station pairs, each telling us the average velocity along a path, we can perform a statistical inversion to create a detailed map of the subsurface. For a given frequency, we have many measurements of travel time $\tau$ over various distances $L$. By plotting $\tau$ versus $L$, we expect to see a straight line whose slope is the inverse of the velocity. By finding the best-fit line through all our data points, we can obtain a robust estimate of the velocity for the region our paths sample. Repeating this for all our frequencies, we build a 3D model of the Earth's crust and upper mantle, revealing ancient tectonic sutures, magma bodies beneath volcanoes, and sedimentary basins.

Perhaps the most revolutionary application of ambient noise interferometry is its ability to monitor the Earth not just in space, but through time. Earth is not a static object; it is a living, breathing body. Magma shifts beneath volcanoes, stress builds and releases along fault lines, and groundwater tables rise and fall. These processes, though often slow and subtle, cause tiny changes in the seismic velocity of the rocks. Before interferometry, detecting these minute changes was nearly impossible.

With our technique, we can compute the Green's function between two stations every single day. If the path between them changes, even slightly, the waveform we retrieve will change. Imagine we have a baseline recording, $s_0(t)$, from a "quiet" period. A month later, after some subtle change in the medium, we record a new waveform, $s_1(t)$. If the velocity of the rock has decreased by a tiny fraction, say 0.1%, it means the waves will take 0.1% longer to travel every leg of their journey. The new waveform will look almost identical to the old one, but it will be slightly stretched in time.

This leads to a wonderfully elegant method for measuring velocity changes, often called the "stretching" method. We take the new waveform $s_1(t)$ and we stretch and squeeze it in time by a variable amount $\epsilon$, creating a family of signals $s_1(t(1+\epsilon))$. We then find the value of $\epsilon$ that makes the stretched signal match the original baseline signal most perfectly. This optimal stretch value, $\epsilon^\star$, directly gives us the fractional velocity change, $\delta v/v = -\epsilon^\star$. Amazingly, the most sensitive part of the signal for this measurement is not the clean, direct arrival, but the long, messy tail of the signal known as the coda. This coda is made of waves that have scattered many times on their journey, sampling the medium far and wide. Because they have traveled for so long, their paths are greatly elongated, so they accumulate a much larger and more easily measurable time shift, making them exquisitely sensitive to tiny, uniform velocity changes.

This technique has transformed volcanology, allowing scientists to watch magma chambers inflate and deflate. It is used to monitor fault zones for changes in stress that might precede an earthquake. On a smaller scale, it is even being used in civil engineering to monitor the structural health of buildings, bridges, and dams for signs of damage or aging. We are, quite literally, using the Earth's hum as a planetary-scale health monitoring system.

So far, we have talked about measuring travel time, or velocity. But this is like describing a symphony only by its tempo. A wave is characterized by more than just its speed; it also has an amplitude (its loudness) and a phase (its timing). When a wave travels through the Earth, it not only takes time, but it also loses energy—a phenomenon called attenuation. A rock that is fractured or contains fluids will damp a wave's energy more than a solid, cold rock. Furthermore, the wave's energy can be focused or defocused by velocity structures, just as a lens focuses light.

Can our noise-based Green's functions tell us about these properties too? The answer is a resounding yes. By analyzing the full wavefield—both its phase and its amplitude—we can create maps of these other physical properties. A powerful technique known as Helmholtz Tomography does just this. It starts from the fundamental wave equation and shows that the spatial variations in the amplitude and phase of our retrieved wavefield are linked. Specifically, the local "curvature" of the amplitude field tells us about focusing and defocusing, while a combination of amplitude and phase gradients reveals the local attenuation [@problem_spt_id:3575711]. This allows us to create maps of the Earth's "fogginess" or "clarity" to seismic waves, providing clues about temperature, fluid content, and rock type that velocity alone cannot give.

Furthermore, many materials in the Earth, especially in the crust and mantle, are anisotropic—their properties depend on the direction you are looking. Think of the grain in a piece of wood; it's much easier to split it along the grain than against it. Similarly, seismic waves may travel at different speeds or attenuate differently depending on their direction of travel relative to the alignment of minerals or cracks in the rock. By using dense arrays of seismometers that record waves traveling in all directions, and by combining the rich information from ambient noise with the clean, direct paths from active sources like seismic vibrators, we can start to untangle these complex directional properties. Advanced inversion strategies are being developed to create maps not just of a single attenuation parameter ($Q$), but of its full anisotropic character, giving us unprecedented insight into the fabric and stress state of the Earth's lithosphere.

We have been discussing all the things we can do with ambient noise, but we have been coy about where this noise actually comes from. The dominant source of the Earth's "hum" in the frequency band typically used for these studies (periods of a few to a few tens of seconds) is the ocean. As ocean waves slosh around in the deep, they interfere and create pressure waves that push and pull on the seafloor, generating seismic waves that travel all around the globe. This is why a seismometer in the middle of a continent can hear the rumblings of a distant storm in the Pacific.

But this is not the only source. In a spectacular example of turning one person's trash into another's treasure, seismologists have discovered that the high-frequency noise of human activity in cities—cars, trains, construction—can also be used for interferometry!. By placing seismometers in an urban area, we can use the ever-present rumble of traffic to retrieve the Green's function and map the very shallow subsurface. This has enormous potential for geotechnical engineering, landslide hazard assessment, and urban planning. Of course, new sources bring new challenges. Unlike the ocean's hum, which can be reasonably diffuse, traffic noise is often highly directional—think of a busy highway. This directionality can bias our measurements, but if we model it correctly, we can still extract accurate information.

This has brought ambient noise interferometry into a new realm, comparing it directly with traditional shallow-geophysics techniques like Multichannel Analysis of Surface Waves (MASW), which uses an active source like a sledgehammer. Studies comparing the two methods have shown that they can provide remarkably consistent results, while also highlighting the unique biases of each—such as the source directionality for passive noise and near-source effects for active methods. This synergy allows for a more complete and robust characterization of the ground beneath our feet.

As with any real-world measurement, applying ambient noise interferometry is an art as well as a science. The Earth is not the perfect, homogeneous sphere of our simple models. One immediate complication is topography: the Earth has mountains and valleys. A wave traveling over a mountain must cover a longer path than one traveling over a flat plain, and this introduces a delay that can be misinterpreted as a region of slow velocity. Fortunately, if we can characterize the statistical properties of the topography—such as its average roughness and correlation length—we can mathematically predict the average bias and the uncertainty this will introduce into our measurements, allowing us to correct for it.

Furthermore, to test our theories and inversion algorithms, we rely heavily on computer simulations. We build numerical models of the Earth and simulate wave propagation to see if we can reproduce our observations. But these numerical methods themselves are not perfect; they introduce their own subtle errors, such as "numerical dispersion," where the simulated wave speed depends on the grid size of the model. Understanding and benchmarking the behavior of these different computational tools is a crucial, if sometimes overlooked, part of the scientific process.

Finally, the sheer volume of data in modern seismology—continuous recordings from thousands of stations worldwide—has opened the door to another interdisciplinary connection: machine learning. We can now turn the problem around and use data science techniques to ask: what environmental conditions give us the best quality Green's functions? By feeding a machine learning model daily proxies for the noise sources—such as wind speed, ocean wave height, and even human traffic patterns—we can train it to predict the quality of our interferometry result. By then interpreting the model, for example using methods like SHAP analysis, we can gain physical insight into what drives our signal. We might discover, for instance, that high ocean waves are the dominant contributor to a good signal, while high local winds degrade it by creating incoherent noise near the receiver. This helps us to not only understand the noise field better but also to intelligently select the best data for our imaging and monitoring tasks.

From the deep mantle to the city streets, from mapping ancient continents to predicting volcanic eruptions, the simple act of cross-correlating noise has given us a profoundly new and powerful way to engage with our planet. It has transformed seismology from a science of intermittent, violent shocks to one of continuous, subtle listening. By embracing the "noise," we have learned to hear the true, perpetual music of the Earth.