> 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/monthly-search-beta/set-up-monthly-search.md).

# Build a detector

Create a Monthly Search (monitoring) project, label examples, train your detector and test it by reviewing a sample of its detections.

{% hint style="warning" %}
Monthly Search is in beta.
{% endhint %}

A Monthly Search (monitoring) detector is a model trained on examples you label. You label, train, then test it by marking a sample of its detections **Good** or **Wrong**. Then you can run it on monthly imagery.

## Before you start

* Your account has Full Access and Monthly Search turned on. See [Who can use it today](https://docs.earthindex.ai/monthly-search-beta/overview.md#who-can-use-it-today).
* You can find at least 10 clear examples of your target and at least 100 squares that are not it, ideally including look-alikes.
* Training counts as one Deep Search toward your monthly allowance (see [Limits and quotas](https://docs.earthindex.ai/reference/limits-and-quotas.md)).

## Steps

1. Click **Start a new project** and select **Monthly Search** under **Project type**. Enter a **Name**, optionally a **Description**, define the **Area of Interest** and click **Create Project**. See [Start a project](https://docs.earthindex.ai/get-started/starting-a-new-project.md).
2. **Build a detector** has three steps: **Label**, **Train** and **Test**. Stay on **Label**.

   ![The Monthly Search side panel on the Build a detector chapter, with the Label, Train and Test stepper and the label counts.](https://182773137-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FIQB2eToY1Iv2IH6cERjn%2Fuploads%2Fgit-blob-28c68d797836e967d2d763b76ff9c001107eed53%2Flabel-stage.jpg?alt=media)
3. Place labels with the map toolbar. See [Label tiles](https://docs.earthindex.ai/search/label-tiles.md). **Add Training Labels** counts **Positive Labels** and **Negative Labels**, each with a tick or cross against its minimum: "10 minimum" and "100 minimum". Choose negatives that look like your target: see [Why negatives matter](https://docs.earthindex.ai/improve-your-results/why-negatives-matter.md).
4. Optional: under **Quickly Find More Examples with AI**, click **Run Quick Predictions**. It needs one positive label and finds up to 50 similar examples. In **Quick Search Predictions**, hover a tile to confirm or reject it, or click **Confirm all** or **Update**.
5. Click **Continue to train**. Training usually takes about 20 seconds. Click **Cancel training** to stop.
6. When **Training complete** appears, the **Test** step opens. Under **Set sensitivity**, choose **Broad**, **Balanced** (default) or **Precise**, or **Custom** for a **Confidence threshold** slider. See [the values](https://docs.earthindex.ai/monthly-search-beta/overview.md#confidence-and-sensitivity).

   ![The Test stage, where you run the detector test and review a sample of its detections.](https://182773137-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FIQB2eToY1Iv2IH6cERjn%2Fuploads%2Fgit-blob-1c59334b2dd098d40091291387709eaad5bd3558%2Ftest-stage.jpg?alt=media)
7. Click **Run detector test**. It can take a few minutes.
8. Review up to 30 detections in three groups of 10: a "High confidence sample", a "Low confidence sample" and a "Random sample". A strip below the map shows your place, such as "Verification 1–10 of 30".
9. Select a tile to see two panels. The Sentinel-2 annual mosaic ("Training imagery") shows what the detector saw. Mapbox Satellite ("Reference only") is sharper but is not the detector's imagery.
10. Mark each detection **Good** or **Wrong**. **Verification Progress** shows **Reviewed**, **Marked good**, **Marked bad** and **Pass rate so far**.
11. When you finish, the app shows the share correct, such as "70% of reviewed examples were correct.", and "Pass threshold: 60%".
    * 60% or more: "Your detector looks ready". Click **Continue to monthly**.
    * Under 60%: "Your detector may need improvement". Click **Add more labels** to return to **Label**, or **Use validation data and retrain**. Or click **Bypass improvements**, then **Bypass and monitor**, to continue as is.

To skip the review, click **Skip human validation**. The app warns that the model will go unchecked. Click **Go back and review** or **Bypass validation**. Skipping is not recommended.

## What to expect

* A tier beside your score: **Excellent** (90% and above), **Strong** (80–89%), **Ready** (60–79%), **Borderline** (40–59%) or **Needs work** (below 40%).
* Your reviews become labels: **Good** is positive and **Wrong** is negative. **Use validation data and retrain** keeps them.
* **Test summary** shows **Holdout accuracy**: accuracy on squares set aside during training.
* **Full Results** (instead of **Focused Review**) shows every test detection. Use **Filter by Confidence** and the list **Unlabeled detections (120)**. After label edits, **Retrain with your edits (3)** replaces **Continue to monthly**.

## If something goes wrong

| Symptom                                                                                                                                         | Fix                                                                                                                                        |
| ----------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ |
| **Continue to train** is disabled                                                                                                               | Hover it: "Add at least 10 positive and 100 negative labels to continue to training." Add the missing labels.                              |
| **Run Quick Predictions** is disabled                                                                                                           | "Add at least one positive label on the map to enable quick predictions."                                                                  |
| "Training is taking longer than usual. Thanks for your patience."                                                                               | Wait. After "Unable to start training. Please try again.", click **Continue to train**.                                                    |
| A yellow notice says a limit was reached when you train                                                                                         | Training uses your monthly Deep Search allowance.                                                                                          |
| "This detector run has no detections to test yet. Retrain the detector, then try again."                                                        | Add positive examples in **Label** and train again.                                                                                        |
| "Only 18 detections were available, so this test uses fewer than 30 examples."                                                                  | Review those. The pass rate uses them.                                                                                                     |
| **Label** and **Train** are locked                                                                                                              | They lock during a test and a review. Click **Clear test results** or finish the review.                                                   |
| "Unable to run the detector test. Please try again." or "The detector finished, but detection files could not be downloaded. Please try again." | Click **Run detector test** again.                                                                                                         |
| You want a new sample later                                                                                                                     | On **Build a detector**, click **Re-run test**.                                                                                            |
| You want a new detector                                                                                                                         | Before you finish a test, click the trash icon on **Training complete** ("Delete Detector"), then **Delete model**. This cannot be undone. |

## Related

* [Monthly Search overview](https://docs.earthindex.ai/monthly-search-beta/overview.md)
* [Run and compare months](https://docs.earthindex.ai/monthly-search-beta/monthly-monitor.md)
