Continuation, Derivatives and Frequency-Domain Methods
Upward and downward continuation, vertical and horizontal derivatives, analytic signal, tilt, reduction to pole and spectral depths for CSIR NET geophysics.
- 20 min
On this page · 15 sections
- Prerequisites
- The Hook: Seeing the Field from a Different Height
- The Field Above the Sources Is Harmonic
- Upward Continuation
- Downward Continuation
- Derivatives
- Reduction to the Pole and Pseudo-Gravity
- Frequency-Domain Interpretation: Depth from the Spectrum
- Memory Tricks
- Don’t Confuse
- Common Mistakes
- Practice
- Quick Revision Table
- Key Takeaways
- What’s Next
Prerequisites
- Corrections and Regional–Residual Separation: the reduced anomaly is the input to every operation here.
- Continuous and Discrete Signals, Fourier Series and Transforms and Sampling Theorem, FFT, Z-Transform and Convolution: wavenumber, the 2D Fourier transform, aliasing, leakage and windows.
- Laplace’s equation and harmonic functions from Fourier Series, Fourier and Laplace Transforms and BSc electrostatics.
- Green’s Theorem, Gauss’ Law, Continuation and the Equivalent Stratum: the potential-theory basis of continuation and equivalent sources.
The Hook: Seeing the Field from a Different Height
The Field Above the Sources Is Harmonic
In any region without sources, the gravitational potential \( U \) and the magnetic scalar potential satisfy Laplace’s equation:
$$\nabla^2 U = \frac{\partial^2 U}{\partial x^2} + \frac{\partial^2 U}{\partial y^2} + \frac{\partial^2 U}{\partial z^2} = 0$$Any derivative of a harmonic function is harmonic, so \( \Delta g = \partial U/\partial z \) and the magnetic field components also satisfy it.
Take the 2D Fourier transform over \( x, y \) with wavenumbers \( k_x, k_y \) and \( \lvert k\rvert = \sqrt{k_x^2 + k_y^2} \). Laplace’s equation becomes
$$\frac{\partial^2 \tilde{U}}{\partial z^2} = \lvert k\rvert^2\,\tilde{U} \quad\Rightarrow\quad \tilde{U}(k, z) \propto e^{\pm\lvert k\rvert z}$$Above the sources the field must decay with height, so only one exponential survives. Taking \( z \) positive downward and observation at height \( \Delta z \) above the measurement plane:
$$\boxed{\tilde{F}(k, z_0 - \Delta z) = \tilde{F}(k, z_0)\,e^{-\lvert k\rvert\,\Delta z}}$$Here \( \lvert k\rvert = 2\pi/\lambda \) in radians per unit length. Every operation on this page is a multiplication of \( \tilde{F} \) by a wavenumber filter; the field at another height, its derivatives, its pole-reduced form are all such products.
Upward Continuation
In the space domain
For a field known over the whole plane \( z = 0 \), the field at height \( h \) above it is a weighted average (the Poisson integral):
$$\boxed{F(x, y, -h) = \frac{h}{2\pi}\iint \frac{F(x', y', 0)}{\left[(x - x')^2 + (y - y')^2 + h^2\right]^{3/2}}\,dx'\,dy'}$$and for a 2D (profile) field
$$F(x, -h) = \frac{h}{\pi}\int_{-\infty}^{\infty}\frac{F(x', 0)}{(x - x')^2 + h^2}\,dx'$$Both kernels integrate to 1: upward continuation is a smoothing average whose width grows with \( h \).
In the wavenumber domain
$$\boxed{\text{Upward continuation filter} = e^{-\lvert k\rvert h} = e^{-2\pi h/\lambda}}$$| Wavelength \( \lambda \) | Factor after continuing up \( h = 1 \) km |
|---|---|
| 1 km | 0.0019 |
| 2 km | 0.043 |
| 5 km | 0.285 |
| 10 km | 0.533 |
| 20 km | 0.730 |
Uses of upward continuation:
- Regional field for regional–residual separation (see the previous page).
- Matching surveys flown at different heights, and comparing ground with airborne data.
- Suppressing shallow noise (cultural sources, near-surface debris).
- Assessing depth: anomalies that persist on strong continuation come from deep or large sources.
Upward continuation is stable: it multiplies every wavenumber by a number less than 1, so noise is reduced.
Downward Continuation
$$\boxed{\text{Downward continuation filter} = e^{+\lvert k\rvert h}}$$Downward continuation brings the observation surface closer to the sources, sharpening anomalies and separating overlapping ones. But it amplifies each wavenumber by \( e^{kh} \), and short wavelengths, where noise lives, are amplified most. Continuing down by 0.5 km multiplies a 1 km wavelength by \( e^{\pi} \approx 23 \), and a 0.5 km wavelength by \( e^{2\pi} \approx 535 \).
Continuation from an uneven surface (a draped survey) to a level one is done with an equivalent source layer: fit a layer of point sources that reproduces the data, then compute their field on the new surface. The equivalent-layer idea follows from Green’s theorem: any harmonic field can be produced by a suitable surface distribution of sources below it.
Derivatives
Vertical and horizontal derivatives in the wavenumber domain
| Operation | Wavenumber filter | Effect |
|---|---|---|
| \( n \)th vertical derivative \( \partial^n F/\partial z^n \) | \( \lvert k\rvert^{n} \) | high-pass; sharpens anomalies |
| Horizontal derivative \( \partial F/\partial x \) | \( i k_x \) | high-pass, directional; peaks over edges |
| Upward continuation by \( h \) | \( e^{-\lvert k\rvert h} \) | low-pass |
| Downward continuation by \( h \) | \( e^{+\lvert k\rvert h} \) | unstable high-pass |
The vertical derivative is never measured on the ground grid; it is computed from the field using Laplace’s equation, which ties it to the horizontal derivatives:
$$\boxed{\frac{\partial^2 F}{\partial z^2} = -\left(\frac{\partial^2 F}{\partial x^2} + \frac{\partial^2 F}{\partial y^2}\right)}$$On a square grid of spacing \( s \), the five-point Laplacian gives a quick second vertical derivative (2VD) at a node with value \( F_0 \) and neighbours \( F_1 \ldots F_4 \):
$$\frac{\partial^2 F}{\partial z^2} \approx \frac{4F_0 - (F_1 + F_2 + F_3 + F_4)}{s^2}$$What derivative maps do:
- Sharpen anomalies and separate adjacent ones; suppress the regional (the regional is broad, so its curvature is small). The 2VD is often used as a residual map.
- Outline bodies: for steep-sided bodies the zero contour of the 2VD lies roughly over the edges; maxima of the horizontal gradient lie over vertical contacts.
- Amplify noise: the filter \( \lvert k\rvert^2 \) boosts short wavelengths quadratically. Derivative maps of noisy data are usually combined with mild upward continuation.
Horizontal gradient magnitude, analytic signal and tilt
For a field \( T \) (gravity, or magnetic data reduced to the pole or transformed to pseudo-gravity):
| Quantity | Definition | Where it peaks or crosses zero | Note |
|---|---|---|---|
| Horizontal gradient magnitude (HGM) | \( \sqrt{(\partial T/\partial x)^2 + (\partial T/\partial y)^2} \) | maxima over vertical contacts | best on gravity or pseudo-gravity |
| Analytic signal amplitude (total gradient) | \( \lvert A\rvert = \sqrt{(\partial T/\partial x)^2 + (\partial T/\partial y)^2 + (\partial T/\partial z)^2} \) | maxima over edges and thin bodies | in 2D independent of magnetisation direction; in 3D only weakly dependent; useful at low latitudes |
| Tilt angle | \( \theta = \tan^{-1}\left[\dfrac{\partial T/\partial z}{\sqrt{(\partial T/\partial x)^2 + (\partial T/\partial y)^2}}\right] \) | zero over vertical contacts; ranges between \( \pm 90° \) | normalises amplitudes, so weak and strong anomalies show equally |
Depth from widths (2D, magnetic):
- Contact: \( \lvert A(x)\rvert \propto 1/\sqrt{x^2 + h^2} \), so the full width at half maximum is \( 2\sqrt{3}\,h \approx 3.46h \).
- Thin dyke: \( \lvert A(x)\rvert \propto 1/(x^2 + h^2) \), so the full width at half maximum is \( 2h \).
- Tilt over a vertical contact (pole-reduced): \( \theta = \tan^{-1}(x/h) \), so the \( \pm 45° \) contours are \( 2h \) apart: the distance from the zero contour to the 45° contour equals the depth to the top.
Reduction to the Pole and Pseudo-Gravity
A total-field anomaly over a body depends on two directions: the inclination and declination of the Earth’s field and of the body’s magnetisation. At the magnetic pole, with vertical field and vertical (induced) magnetisation, the anomaly of a symmetric body is a symmetric high directly over it. Elsewhere it is skewed, with a high and a low (a dipolar pair), and its peak is offset from the body.
Reduction to the pole (RTP) transforms the anomaly to what it would be at the pole. In the wavenumber domain it divides by two direction factors:
$$\boxed{\tilde{T}_{RTP} = \frac{\tilde{T}}{\Theta_m\,\Theta_f}} \qquad \Theta = \hat{n}_z + i\,\frac{\hat{n}_x k_x + \hat{n}_y k_y}{\lvert k\rvert}$$where \( \hat{\mathbf{n}} \) is the unit vector of the magnetisation (\( \Theta_m \)) or field (\( \Theta_f \)), with \( \hat{n}_z = \sin I \).
- At the pole \( \hat{n}_z = 1 \), \( \hat{n}_x = \hat{n}_y = 0 \): \( \Theta = 1 \), no change.
- At low magnetic latitudes (\( I \) small), for wavenumbers perpendicular to the declination the factor \( \Theta \) is close to \( \sin I \), and the filter amplifies those wavenumbers by about \( 1/\sin^2 I \). The RTP becomes unstable and produces strong artefacts elongated along the declination. Remedies: reduction to the equator, stabilised (amplitude-limited) RTP, or use of the analytic signal.
- RTP assumes the magnetisation direction is known, usually that it is induced (parallel to the present field). Strong remanence in another direction makes the RTP wrong.
Pseudo-gravity. Poisson’s relation states that for a body with uniform density and uniform magnetisation, the magnetic potential is proportional to the derivative of the gravitational potential in the direction of magnetisation. So a magnetic anomaly can be transformed into the gravity anomaly the body would produce if its magnetisation were replaced by a proportional density, the pseudo-gravity: RTP followed by vertical integration (a filter \( 1/\lvert k\rvert \)). Pseudo-gravity is compared directly with real gravity to test whether the same bodies cause both; its horizontal gradient maxima locate magnetic contacts.
Frequency-Domain Interpretation: Depth from the Spectrum
The radially averaged power spectrum
Compute the 2D power spectrum \( P(k_x, k_y) = \lvert\tilde{T}\rvert^2 \) of a gridded anomaly and average it in rings of constant \( \lvert k\rvert \). For an ensemble of sources whose tops lie at a mean depth \( h \) below the observation plane, the depth term dominates and
$$\boxed{\ln P(k) \approx C - 2h\,k}$$with \( k \) in radians per unit length. The slope of \( \ln P \) against \( k \) is \( -2h \). If \( k \) is in cycles per unit length (\( f = k/2\pi \)), the slope is \( -4\pi h \).
Real spectra usually show two or three straight segments: a steep low-wavenumber segment from deep sources (for example basement), a gentler segment from shallow sources, and a flat high-wavenumber tail of noise. Each segment’s slope gives its ensemble’s mean depth; the crossover between segments suggests a cut-off for a matched filter that separates the two.
The statistical model behind this, an ensemble of prismatic sources with random properties, also has a size factor and a thickness factor: the bottom of the sources flattens the spectrum at low wavenumber and can produce a spectral peak. That is used to estimate the depth to the bottom of magnetic sources, interpreted as the Curie-point depth where crustal rocks lose ferromagnetism (magnetite’s Curie temperature is about 580 °C). In the centroid approach, the depth to the top \( z_t \) comes from the high-wavenumber slope and the centroid depth \( z_0 \) from the low-wavenumber part; the bottom is
$$z_b = 2z_0 - z_t$$Curie depths are compared with heat flow: a shallow Curie depth suggests a high geothermal gradient.
Other wavenumber-domain filters
| Filter | Purpose |
|---|---|
| Low-pass, high-pass, band-pass (Butterworth, cosine-tapered) | separate wavelength bands; taper to avoid ringing (see Filters, Deconvolution and Fractal Analysis) |
| Matched filters | designed from the spectral segments to pass one source ensemble |
| Directional (strike) filters | remove or enhance features of one orientation, such as flight-line noise |
| Wiener filters | optimal separation of signal and noise with known spectra |
Practical cautions for Fourier processing
- Edges: the FFT assumes the grid repeats periodically; a jump between opposite edges causes leakage and edge artefacts. Remove a trend, extend the grid smoothly (padding) and taper before transforming.
- Grid spacing: the Nyquist wavenumber is \( \pi/\Delta x \); anomalies narrower than two grid cells are aliased.
- Grid size: the longest wavelength resolved is about the grid width; deep sources need large areas.
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Upward continuation: stable low-pass, physically exact | low-pass filter with arbitrary cut-off: not tied to a height | |
| Downward continuation: valid only above the sources | continuation through sources: meaningless | |
| Vertical derivative filter \( \lvert k\rvert \) | horizontal derivative filter \( ik_x \): directional, phase-shifted | |
| Horizontal gradient maxima over contacts | tilt-angle zero over contacts | |
| Analytic signal in 2D: independent of magnetisation direction | RTP: requires the magnetisation direction | |
| Slope \( -2h \) with \( k \) in rad/length | slope \( -4\pi h \) with \( f \) in cycles/length | |
| Depth to top (high-wavenumber slope) | depth to bottom / Curie depth (\( 2z_0 - z_t \)) | |
| Pseudo-gravity: magnetic data turned into equivalent gravity | pseudo-section: an electrical-method display (unrelated) |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper.
Level 1: Foundation
- Write the wavenumber-domain filter for upward continuation by a height \( h \).
- Which equation allows the vertical derivative to be computed from horizontal data?
- Over a vertical contact (pole-reduced data), what is the value of the tilt angle?
Answers
- \( e^{-\lvert k\rvert h} \), with \( \lvert k\rvert = 2\pi/\lambda \).
- Laplace’s equation: \( \partial^2F/\partial z^2 = -(\partial^2F/\partial x^2 + \partial^2F/\partial y^2) \).
- Zero.
Level 2: Part B Style
Practice Q1. A 4 km wavelength component of a gravity anomaly is upward-continued by 2 km. Its amplitude is multiplied by about:
(a) 0.61 (b) 0.21 (c) 0.043 (d) 0.0019
Answer: (c) 0.043
\( e^{-2\pi h/\lambda} = e^{-2\pi\times 2/4} = e^{-\pi} = 0.0432 \).
The distractors are slips in the exponent: (b) is \( e^{-\pi/2} = 0.208 \) (height taken as 1 km instead of 2 km); (a) is \( e^{-h/\lambda} = e^{-0.5} = 0.61 \) (forgetting \( 2\pi \)); (d) is \( e^{-2\pi} = 0.0019 \) (wavelength taken as 2 km).
Practice Q2. Downward continuation of potential-field data is called unstable because it:
(a) attenuates long wavelengths (b) amplifies short-wavelength components, including noise, exponentially (c) changes the sign of the anomaly (d) violates Laplace’s equation above the sources
Answer: (b)
The filter \( e^{+kh} \) grows with wavenumber. Above the sources downward continuation is mathematically valid (so (d) is wrong); below the top of the sources it fails.
Practice Q3. The slope of a radially averaged log power spectrum, \( \ln P \) plotted against wavenumber in rad km\(^{-1}\), is \( -5.0 \) km for the deep segment. The mean depth to the tops of these sources is:
(a) 0.40 km (b) 2.5 km (c) 5.0 km (d) 10 km
Answer: (b) 2.5 km
\( \ln P = C - 2hk \), so \( h = 5.0/2 = 2.5 \) km. Option (c) uses the amplitude-spectrum relation; (a) divides by \( 4\pi \) as if \( k \) were in cycles per km.
Practice Q4. Which transformation of magnetic data is most likely to produce strong artefacts in a survey near the magnetic equator?
(a) Upward continuation (b) Analytic signal amplitude (c) Reduction to the pole (d) Horizontal derivative
Answer: (c)
The RTP filter amplifies wavenumbers across the declination by roughly \( 1/\sin^2 I \), which diverges as \( I \to 0 \).
Level 3: Part C Style
Practice Q5. A 2D profile contains two components: \( \lambda_1 = 10 \) km, amplitude 8 mGal, and \( \lambda_2 = 1 \) km, amplitude 1 mGal. Consider:
- After upward continuation by 0.5 km the 1 km component is reduced to about 0.04 mGal.
- In the first vertical derivative, the 1 km component has a larger amplitude than the 10 km component.
- Downward continuation by 0.5 km multiplies the 10 km component by about 1.37.
- Upward continuation changes the wavelengths of both components.
Which are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 2, 3 and 4 only (d) 1, 2, 3 and 4
Answer: (b) 1, 2 and 3 only
- \( 1\times e^{-2\pi\times 0.5/1} = e^{-\pi} = 0.043 \) mGal. Correct.
- Derivative amplitudes: \( 8\times 2\pi/10 = 5.03 \) and \( 1\times 2\pi/1 = 6.28 \) mGal km\(^{-1}\). Correct.
- \( e^{+2\pi\times 0.5/10} = e^{0.314} = 1.37 \). Correct.
- Wrong. A linear filter changes amplitude (and for some filters phase), never the wavelength of a sinusoid.
Practice Q6. A radially averaged power spectrum of an aeromagnetic grid flown 300 m above the ground, plotted as \( \ln P \) against \( f \) in cycles per km, shows two segments: from \( f = 0.05 \) to 0.25 the value falls from 12.0 to 2.0; from \( f = 0.25 \) to 0.75 it falls from 2.0 to −1.8. The depths of the two source ensembles below ground are closest to:
(a) 4.0 km and 0.60 km (b) 3.7 km and 0.30 km (c) 25 km and 3.8 km (d) 3.7 km and 0.60 km
Answer: (b) 3.7 km and 0.30 km
With \( f \) in cycles per km, slope \( = -4\pi h \).
Deep: slope \( = (2.0 - 12.0)/(0.25 - 0.05) = -50 \) km, so \( h = 50/4\pi = 3.98 \) km below the sensor. Below ground: \( 3.98 - 0.30 = 3.68 \approx 3.7 \) km.
Shallow: slope \( = (-1.8 - 2.0)/(0.75 - 0.25) = -7.6 \) km, so \( h = 7.6/4\pi = 0.605 \) km below the sensor, or \( 0.605 - 0.30 = 0.30 \) km below ground.
Option (a) forgets to subtract the flight height; (c) uses \( h = -\text{slope}/2 \), treating \( f \) as radian wavenumber; (d) subtracts the flight height for one ensemble only.
Practice Q7. The analytic-signal amplitude of a 2D magnetic anomaly has a single peak with full width at half maximum 1.4 km. If the source is a thin vertical dyke, its depth to top is \( h_1 \); if it is a vertical contact, \( h_2 \). Then \( (h_1, h_2) \) is closest to:
(a) (0.70 km, 0.40 km) (b) (1.4 km, 0.40 km) (c) (0.70 km, 1.21 km) (d) (0.35 km, 0.70 km)
Answer: (a) (0.70 km, 0.40 km)
Dyke: FWHM \( = 2h \), so \( h_1 = 0.70 \) km. Contact: FWHM \( = 2\sqrt{3}h \), so \( h_2 = 1.4/3.464 = 0.404 \) km.
The interpretation depends on the assumed source type, which is why the structural index matters in automatic depth methods (next page).
Quick Revision Table
| Idea | Formula or fact |
|---|---|
| Harmonic field | \( \nabla^2 F = 0 \) above sources; \( \tilde{F} \propto e^{-\lvert k\rvert z} \) |
| Upward continuation | \( e^{-\lvert k\rvert h} \), stable low-pass; Poisson integral kernel |
| Downward continuation | \( e^{+\lvert k\rvert h} \), unstable, only above sources |
| \( n \)th vertical derivative | \( \lvert k\rvert^n \) |
| Horizontal derivative | \( ik_x \) |
| 2VD on a grid | \( [4F_0 - \sum F_i]/s^2 \) |
| HGM | maxima over vertical contacts |
| Analytic signal | \( \sqrt{T_x^2 + T_y^2 + T_z^2} \); FWHM contact \( 2\sqrt3h \), dyke \( 2h \) |
| Tilt | \( \tan^{-1}(T_z/\lvert\nabla_hT\rvert) \); zero at contact, \( \pm45° \) at \( \pm h \) |
| RTP | divide by \( \Theta_m\Theta_f \); unstable at low inclination |
| Pseudo-gravity | Poisson’s relation; RTP + \( 1/\lvert k\rvert \) |
| Point mass spectrum | \( 2\pi GM e^{-\lvert k\rvert z} \) |
| Spectral depth | \( \ln P = C - 2hk \) (rad); \( -4\pi h \) (cycles) |
| Curie depth | \( z_b = 2z_0 - z_t \) |
Key Takeaways
What’s Next
- Interpretation of Simple Bodies, Forward Modelling and Inversion: sphere, cylinder, slab and dyke formulas, depth rules, Euler deconvolution, Talwani modelling and inversion.
- Gravity and Magnetic Methods: Quick Revision.
- Signal-processing background: Filters, Deconvolution and Fractal Analysis.