- Best estimate
- 44.0 km
- Uncertainty σ
- 18.0 km
Lab · Exploration 13
The Next Observation
A storm is approaching the coast. An aircraft has only a few instruments. Where should the next one be used?
6 dropsondes available
The best next observation is not automatically where uncertainty is largest. It is the observation most likely to change what we know about the target that matters.
More data is not automatically better evidence
A measurement can be precise and still have little effect on the forecast we care about. A highly uncertain region can be dynamically disconnected from the future target. A sensitive region may already be well observed. Two instruments may repeat the same information, while a different measurement reveals something the entire network has missed.
Where do we know the least?is not the same question as
What measurement could change what we know about the outcome that matters?
Watch first · short video
More Data ≠ More Understanding
Begin with a crucial distinction: the next available measurement is not always the next useful piece of evidence.
Watch video ↗Applied Mathematics in Geosciences · Episode 1
Then explore it yourself ↓
Choose the question before the instrument
There is no universally best observing location because there is no universally best target. The same atmosphere supports several legitimate questions, and each gives a different value to the same measurement.
Selected target
48-hour landfall
The atmosphere has not changed. The question has. A measurement near the storm center is valuable for intensity, while an upstream wind measurement may be more valuable for landfall.
- Quantity
- Landfall location
- Region
- Central coastline
- Lead time
- 48 hours
- Goal
- Reduce uncertainty in this outcome
Where should the next observation go?
We want to reduce uncertainty in the selected future target. Choose one location, receive one synthetic measurement, and compare the target before and after.
Choose one observation location
The candidate samples part of the present state. After it arrives, the present state is updated and the ensemble is propagated to the 48-hour target.
In the real atmosphere the exact truth is unknown. Here a synthetic truth lets us see whether the observation actually improved reconstruction and forecast.Candidate C→waiting
Choose a candidate observation
Why uncertainty, sensitivity, and expected value differ
Current uncertainty asks where plausible present states differ. Target sensitivity asks which differences can reach the future target. Expected value combines that connection with existing coverage and measurement noise.
Δj(D)=Var(Z | D) − Var(Z | D, Yj)
More precision or a new direction?
An observing network can be precise and still miss an important part of the state. Repeating the same measurement reduces noise, but may not remove ambiguity about hidden structure.
Add an observation
Choose an addition to the observing network.
What remains possible?
Surface data narrow the band but leave it long along the hidden direction.
Complementary observations reveal different directions. Redundant observations mostly repeat what is already known.
Repeated measurements can still detect bias, provide backup, preserve long records, and average down random error. The point is that greater precision cannot reveal a direction the observing system never measures.
A good observation can test an explanation
Sometimes the goal is not merely to narrow one forecast. It is to distinguish explanations that both reproduce the familiar data.
Subsurface-memory model
Heat stored below the surface and thermocline changes create a delayed ocean response.
Strong surface-feedback model
A stronger direct surface feedback compensates for weaker subsurface memory.
Agreement with the same visible pattern does not guarantee agreement about the mechanism.
Model AModel B
Stronger support for Model A
The observation falls nearer the subsurface-memory prediction.
How different are the measurable predictions?
Sj=|m1j − m2j|√(q1j2 + q2j2)
The numerator is the distance between predicted centers. The denominator is the blur caused by predictive uncertainty and measurement error. A large score means a real instrument may be able to distinguish the predictions.
Observe where plausible explanations make measurably different predictions, not only where they already agree.
After one observation, what should come next?
Choose one at a time. Each measurement changes the ensemble, so the value of every remaining candidate must be recomputed.
Observation 1 of 3
3 remainingReduce uncertainty in 48-hour landfall
- Best estimate
- 51.0 km
- Uncertainty σ
- 18.0 km
Candidate C→next result
- Updated estimate
- 52.0 km
- Uncertainty σ
- 12.8 km
Observation value changed after the update. The first result changed what is still uncertain, so the next choice is not governed by a permanent ranking.
Add real-world constraints
The mathematically highest-value observation is not useful if it cannot be collected safely, accurately, and in time.
observation value changes after every update
Final synthesis
The next observation is a designed question
A useful observation is not defined only by where uncertainty is largest. Its value depends on the question, the dynamics connecting present and future, the observations already available, instrument limitations, and the explanations the evidence could test.
- What target matters?
- What is still uncertain?
- Can information from this location reach the target?
- Does the current network already supply the same information?
- Would the measurement distinguish plausible explanations?
- Can it be collected safely, accurately, and in time?
- Should the next choice be reconsidered after this evidence arrives?
Do not collect the next measurement simply because it is available. Choose it because its possible answers could change what we know.
A carefully chosen observation can sharpen a forecast or challenge an explanation. It still cannot prove that a model is true.
Sources, methods, and synthetic-data note
- Literature on targeted observations, adaptive sampling, data assimilation, and expected reduction in forecast uncertainty.
- Linear–Gaussian observation-value calculations and elementary model-discrimination scores.
- ENSO-inspired conceptual models used only to illustrate observational tests of competing mechanisms.
Every storm track, map, ocean state, candidate measurement, and model response on this page is generated from lightweight synthetic teaching data. Rankings are conditional on the selected target and displayed assumptions; none is operational guidance.