Skip to content

Precision, Recall, and Classification Intuition

Why These Terms Feel Confusing

Different communities often use different vocabulary for the same underlying ideas.

  • Machine learning talks about:
  • false positives
  • false negatives
  • precision

  • recall

  • Remote sensing often talks about:

  • omission error
  • commission error
  • producer's accuracy
  • user's accuracy

The goal of this note is to build a simple conceptual understanding that remains intuitive even months later.

Start with the Main Idea

Suppose we are interested in detecting something:

  • water
  • cancer
  • flood extent
  • glacier pixels
  • spam email
  • defects in manufacturing

The thing we are interested in detecting is called the positive class.

Everything else is the negative class.

For example:

Positive  = water
Negative  = non-water

or

Positive  = cancer
Negative  = healthy

This simple framing makes the rest of the terminology much easier.

Core Classification Outcomes

Mental Model

A classifier compares:

REALITY  --->  MODEL PREDICTION

The prediction can either agree or disagree with reality.

The Four Possible Outcomes

Reality Prediction ML Term Remote Sensing Term Intuition
Positive Positive True Positive (TP) Correct detection Correctly found the phenomenon
Positive Negative False Negative (FN) Omission error Missed phenomenon / falsely predicted as negative
Negative Positive False Positive (FP) Commission error False alarm / false inclusion
Negative Negative True Negative (TN) Correct rejection Correctly ignored non-target

False Negative (FN): The Most Important Intuition

A false negative occurs when:

Something truly positive
is falsely predicted as negative.

Example:

Reality:     Water
Prediction:  Non-water

The water existed. The classifier missed it.

This is:

  • a False Negative (FN)
  • an omission error
  • a missed detection

Cancer Example

Reality:     Cancer exists
Prediction:  Healthy

The disease existed, but the detector failed to identify it.

False Positive (FP)

A false positive occurs when:

Something truly negative
is falsely predicted as positive.

Example:

Reality:     Road or shadow
Prediction:  Water

This is:

  • a False Positive (FP)
  • a commission error
  • a false alarm

Recall and Precision

Concept Question Equation Core Intuition
Recall Out of all real positive cases, how many did we recover? Recall = TP / (TP + FN) Completeness / sensitivity
Precision Out of everything predicted as positive, how many were actually positive? Precision = TP / (TP + FP) Purity / trustworthiness

Notice that both metrics share the same numerator:

TP = correctly detected positive cases

The denominator changes:

  • Recall is anchored to reality.
  • Precision is anchored to model predictions.

Why Not Just Use Accuracy?

Accuracy is the most familiar metric:

Accuracy = (TP + TN) / (TP + TN + FP + FN)

It answers a simple question: out of everything, how much did the classifier get right?

That question becomes misleading the moment the positive class is rare, which is the normal case in remote sensing and many real-world detection problems.

Suppose flood pixels make up 1% of an image. A classifier that predicts:

Everything is non-flood.

gets every non-flood pixel right and every flood pixel wrong. Yet:

Accuracy = 99%

The classifier never once detected the actual phenomenon of interest, and the headline number still looks excellent.

This is the same failure mode as the earlier "everything is water" example, just measured with the wrong metric:

  • Recall = 0% (no flood pixels recovered) — correctly exposes the failure
  • Accuracy = 99% — hides the failure entirely

The rarer the positive class, the more accuracy rewards a model for ignoring it. This is exactly why:

  • cancer screening,
  • flood mapping,
  • glacier/ice detection,
  • fraud and defect detection,

are never evaluated with plain accuracy. In each of these, the positive class is rare relative to the negative class, so accuracy stays high even when the model is essentially non-functional. Precision and recall are anchored to the positive class specifically (TP, FN, FP), which is what makes them resistant to this failure mode.

Rule of thumb: whenever the phenomenon of interest is rare compared to everything else in the dataset, accuracy alone is not a trustworthy metric. Reach for recall, precision, and F1 instead.

Flood Mapping Example

Suppose we are mapping flood extent after a hurricane.

Missing flooded areas can be dangerous:

  • rescue teams may miss affected regions
  • risk estimates become too low
  • emergency response becomes incomplete

In this situation, high recall is often preferred.

Extreme Recall Example

Suppose a classifier predicts:

Everything is water.

Then:

  • no water pixels are missed
  • recall becomes very high (possibly 100%)

But many non-water pixels are also incorrectly classified as water.

So:

  • high recall alone is not enough

Medical Screening Intuition

In medicine, missing a dangerous disease is often considered worse than temporarily flagging healthy individuals for additional testing.

During the early stages of the COVID-19 pandemic, many public-health strategies emphasized avoiding missed cases, illustrating a high-recall mindset.

Spam Email Example

Suppose a spam filter incorrectly sends important emails into the spam folder.

Users quickly lose trust in the system.

In this case:

  • high precision matters strongly
  • emails classified as spam should truly be spam

Extreme Precision Example

Suppose a model predicts only one water pixel in an entire image.

If that single pixel is truly water:

Precision = 100%

But almost all real water pixels were missed.

So:

Recall ≈ 0

This shows:

  • precision alone is also not enough

Why We Need Both Precision and Recall

Goal Metric
Capture as much real phenomenon as possible Recall
Ensure detections are trustworthy Precision

A good classifier usually needs balance between both.

F1 Score

The F1 score combines precision and recall into a single metric.

F1 = 2 * (Precision * Recall) / (Precision + Recall)

The F1 score becomes low when either:

  • precision is poor
  • recall is poor

This prevents "cheating" by:

  • predicting everything positive
  • or predicting almost nothing positive

Conceptually:

F1 rewards models that are both:
- complete
- trustworthy

Vocabulary Bridge

Scientific / Remote Sensing Language ML / Statistics Language Core Meaning
Omission error False Negative (FN) Missed positive phenomenon
Commission error False Positive (FP) False inclusion / false alarm
Producer's accuracy Recall How much real phenomenon was captured
User's accuracy Precision How trustworthy detections are
Completeness Recall Detection coverage
Purity Precision Detection quality

Mental Model Summary

Recall

Did we capture the real phenomenon?

Recall penalizes:

  • missed detections
  • omission errors
  • false negatives

Precision

Can we trust our predictions?

Precision penalizes:

  • false alarms
  • commission errors
  • false positives

Final Conceptual Takeaway

A classifier is fundamentally making decisions about:

what belongs to the phenomenon of interest
and what does not.

Once the positive class is clearly defined:

  • false negatives become missed positives
  • false positives become false alarms
  • recall measures completeness
  • precision measures trustworthiness

Understanding these concepts intuitively is far more valuable than memorizing formulas.