Ocean layer
Many vertical temperature profiles can have the same average.
Module II begins
Module I followed how a state evolves and why its future can become difficult to predict. Module II begins with an earlier question: how do we know the present state in the first place?
Module II · Reconstructing the unseen
Can you reconstruct a world you cannot observe directly?
Earth-system maps often look complete, but the measurements behind them are scattered, indirect, and uncertain. In this lab, you will build a hidden field from sparse observations, discover several worlds that fit the same data, and test which details the observing system can actually recover.
Exploration 6 · 12–18 minute explorationPlaying
Watch first · short video
Why Models Need Data—and Why Data Need Models
Begin with the two-way relationship between incomplete observations and the models used to reconstruct what lies between them.
Watch video ↗Applied Mathematics in Geosciences · Episode 10
Related: More Data ≠ More Understanding ↗Then explore it yourself ↓
A measurement is not the state
The hidden field may contain many values, but an instrument returns only one particular view of that field. What it reports depends on how the instrument samples the world.
y = H(x) + η
hidden state (x)the field we want to know→measurement rule (H)what the instrument responds to→reported observation (y)the number we receive
Changing the instrument changes H, even when the Earth state x is exactly the same.
Try this: keep the hidden field fixed and switch between the three instruments.Watch the reported number change even though the hidden Earth state has not changed.
Hidden field
Why this mattersA measurement is not a miniature copy of the hidden state. It is the result of a measurement rule. To interpret an observation, we must know what the instrument actually responds to.
same Earth state + different H ⇒ different observations
Several hidden worlds can fit the same measurements
The measurements below are exact. What do you think happens between them? Draw a possible field or choose one, then compare it with alternatives that fit the same evidence.
Precision reduces uncertainty in what the instrument already measures. A new observation can reveal a part of the state that was previously unconstrained.
Seismic path
Different arrangements of wave speed can produce similar arrival times.
Satellite pixel
Different mixtures of cloud, land, and water can produce a similar signal.
A reconstruction needs more than data
When observations do not determine a unique answer, a reconstruction must prefer some possibilities over others. That preference may come from smoothness, physics, historical behavior, or a reference state.
Follow every measurementBalance data and structurePrefer a smoother field
Why this mattersRegularization makes an assumption visible and adjustable. A weak assumption may allow noise to dominate. A strong assumption may suppress real structure. The best choice cannot be judged from appearance alone.
Show the reconstruction score
reconstruction score = data mismatch + regularization strength × structure penalty
Data mismatch asks how closely the reconstruction agrees with the measurements.
Regularization strength sets how strongly the structural preference matters.
Structure penalty increases when the curve contains the kind of roughness this reconstruction is designed to avoid.
Can the system tell two nearby features apart?
Scientific resolution is not the size of the pixels on the final map. It is the smallest structure that the complete observing and reconstruction system can reliably distinguish.
Synthetic truth
What the observing system reconstructs
Detected?Yes
Located correctly?Yes
Can it separate the two features?Not yet
display resolution ≠ scientific resolution
Truth → Measurements → Reconstruction → CompareA synthetic resolution test starts with a feature we know, passes it through the measurement process, and reconstructs it using the same method used for unknown fields. We then ask what survived.
What can be lost?Blur, displacement, merging, and amplitude loss reveal the resolution of the complete measurement–reconstruction chain.
Advanced experiment settings
Where would one new observation help most?
Candidate A and Candidate B are both consistent with the observations collected so far. Another measurement is valuable only if the two candidates predict different values there.
distinguishing power = |H(xA) − H(xB)|
A useful location is one where the two candidate worlds make different measurement predictions.Where would you put one sensor?Drag the orange sensor. Look for a location where Candidate A and Candidate B are far apart.
Excellent placement. The two possible worlds make clearly different predictions here.
strong discriminationMore observations are not automatically more informative. The best next observation targets what the current data still cannot distinguish.
Observe where the remaining ambiguity is most exposed.
Mission synthesis
What did the reconstruction mission reveal?
- 1
Measurements are selective.
- 2
Different hidden states can fit the same data.
- 3
Assumptions stabilize the reconstruction.
- 4
Visual detail can exceed evidence-supported detail.
- 5
New observations should target the remaining ambiguity.
A complete-looking Earth map is usually a reconstruction. Its scientific value depends not only on how well it matches the available data, but also on what the instruments can detect, which assumptions fill the gaps, how uncertainty is represented, and whether independent observations can test the result.
For readers who want the mathematics
Linear observation model
y = Hx + η
H maps a hidden state to the measurements an observing system would produce.
Invisible change
If Hv = 0, then adding v changes the hidden state without changing the ideal observations. In linear algebra, such changes form the null space of H.
Regularized reconstruction
minimize: data mismatch + λ × structure penalty
The parameter λ controls the balance between matching observations and preferring the chosen structure.
Resolution
A reconstruction operator maps observations back to an estimated state. The combined observation–reconstruction system determines which features are recovered, blurred, or lost.
Sources, method, and synthetic-data note
- A. Tarantola, Inverse Problem Theory and Methods for Model Parameter Estimation, SIAM.
- P. C. Hansen, Discrete Inverse Problems: Insight and Algorithms, SIAM.
- C. D. Rodgers, Inverse Methods for Atmospheric Sounding, World Scientific.
Every hidden field, measurement, and reconstruction on this page is deterministic synthetic teaching data. The one-dimensional examples combine smooth Gaussian and sinusoidal features. Compatible fields use perturbations that vanish at observed locations. The smoothing experiment uses a regularized Fourier representation, and the two-dimensional recovery tests use Gaussian features sampled and reconstructed from controlled sensor geometries. These demonstrations isolate inverse-problem ideas; they are not Earth observations or operational reconstruction products.