> 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/fix-bad-results.md).

# Fix poor results

Diagnose and fix the two kinds of poor search results, false positives and misses, with a table of causes and fixes.

A search can be wrong in two ways. It can return things that are not your target (false positives). Or it can fail to return things that are (misses). The fix depends on which one you see.

## The short version

* False positives call for negative labels and a higher confidence.
* Misses call for more varied positives and a lower confidence.
* Check your progress by counting how many of the top 50 results are right.

## Find the cause

| What you see                                   | Why                                                  | What to try                                                                                                                                     |
| ---------------------------------------------- | ---------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------- |
| Many look-alikes are returned                  | The search cannot tell them from your target         | Add the look-alikes as negatives. Raise the minimum confidence.                                                                                 |
| Examples you know about are missing            | Your positives do not cover their look               | Label some of them as positives. Check the confidence slider is not too high. Check they are larger than a few 10 m pixels.                     |
| All results are in one place                   | Your labels are clustered                            | Label in several parts of the area.                                                                                                             |
| One project mixes two different looks          | One set of labels cannot describe both               | Split them into separate projects.                                                                                                              |
| Results look random with only 1 or 2 positives | There is too little to learn from                    | Add varied positives.                                                                                                                           |
| Deep Search returns too few or too many        | The confidence threshold or negative sampling is off | Move the **Minimum Confidence** slider. Try **Skip Random Negatives**.                                                                          |
| Monthly counts jump between months             | The seasons change the imagery                       | Compare **Label Transfer** and **Window Calibration**. See [Monthly Search](https://docs.earthindex.ai/monthly-search-beta/monthly-monitor.md). |
| Patches of results sit under clouds or haze    | Clouds change what the satellite sees                | Compare imagery from the same season.                                                                                                           |

## How to tell if you are improving

After each run, look at the top 50 results and count how many are right. If the count goes up from run to run, your labels are getting better.

Monthly Search uses the same idea. Its test step shows you a sample of detections and passes your detector when 60% are correct. Use the same habit with Quick Search and Deep Search, even though they do not run a test for you.

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

Change one thing at a time. Fix the most common wrong result first, run again, and count the top 50 before you change anything else.
{% endhint %}

<details>

<summary>Go deeper</summary>

False positives and misses trade off against each other through the confidence threshold. A higher threshold returns fewer tiles and removes false positives, but it also drops true matches. A lower threshold does the opposite. Labels move the whole trade-off, so they are a better fix than the threshold alone. See [How Earth Index finds matches](https://docs.earthindex.ai/improve-your-results/how-search-works.md).

</details>

## Related

* [Why negatives matter](https://docs.earthindex.ai/improve-your-results/why-negatives-matter.md)
* [Choose good examples](https://docs.earthindex.ai/improve-your-results/choose-good-examples.md)
* [What the satellite sees](https://docs.earthindex.ai/improve-your-results/beyond-true-color.md)
