> For the complete documentation index, see [llms.txt](https://docs.earthindex.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.earthindex.ai/improve-your-results/why-negatives-matter.md).

# Why negatives matter

Why negative labels matter, which negatives help most, and how Quick Search, Deep Search and Monthly Search each use them.

Positive labels say what you want. Negative labels say where it ends. Without negatives, a search cannot tell your target from the things that look like it.

## The short version

* Most of any map is not your target, and "not your target" is varied: forest, fields, water, roofs, shadow.
* The most useful negatives look like your target. An airstrip's best negative is a road.
* Each search mode uses negatives differently, so the number you need differs by mode.

![Two panels of dots. With positives only, the boundary is loose and takes in look-alikes. With negatives added, the boundary is tight around the target.](https://182773137-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FIQB2eToY1Iv2IH6cERjn%2Fuploads%2Fgit-blob-4254f48f690938e5946d88c3e1f0dea4fd3c4b72%2Fnegatives-boundary.svg?alt=media)

## Positives alone cannot say where the target ends

Imagine you label five airstrips. Earth Index learns what airstrips look like, but not what to exclude. Roads, dry riverbeds and cleared fields share some of that look, so they come back as results.

Mistakes cluster on look-alikes. When you label a road as negative, you teach the search the difference between a road and an airstrip. That one label removes a whole group of wrong results.

## Choose negatives that resemble the target

A negative from the middle of a lake does little. A lake is already far from an airstrip, so labelling it changes almost nothing. A negative from a road next to an airstrip does a lot, because it sits right at the edge.

Pick negatives from the wrong results you actually see. Those are the cases the search is confused about.

## How each mode uses negatives

* **Quick Search** moves its search point away from the average of your negatives. Negatives that look like the target help. A heap of unrelated negatives averages out to little.
* **Deep Search** learns a boundary between your positives and your negatives. By default it also adds random negatives from your area. So your own negatives should be the look-alikes. If you tick **Skip Random Negatives**, your negatives are the only ones, so label more and more varied negatives.
* **Monthly Search** adds no negatives for you. This is why the app asks for at least 100 negatives before it trains.

## Collect negatives cheaply

* Run a search, then mark wrong results as negative with Quick Validation.
* Paint a look-alike area with the paintbrush tool to add many negatives at once.

{% hint style="success" %}
**What to do**

After each run, find the three most common wrong results and label one or two of each as negative. Then run again.
{% endhint %}

<details>

<summary>Go deeper</summary>

Deep Search fits a logistic regression to the embeddings of your labelled tiles. It then scores every tile in the area and returns a tile as a detection when its probability is above the **Minimum Confidence** you set. A logistic regression separates two groups with one boundary, so it needs examples from both sides to place that boundary well.

* [Logistic regression (scikit-learn)](https://scikit-learn.org/stable/modules/linear_model.html#logistic-regression)

</details>

## Related

* [Quick Search or Deep Search](https://docs.earthindex.ai/improve-your-results/quick-vs-deep-search.md)
* [Choose good examples](https://docs.earthindex.ai/improve-your-results/choose-good-examples.md)
* [Quick Search](https://docs.earthindex.ai/search/quick-search.md)
