Running predictions

Once you've assigned a few labels, you're ready to run predictions.

Running predictions

To begin, click "Run Predictions".

Soon after, a series of tiles will appear on the Predictions window. These are tiles that are most similar to the positively labelled tiles. On the map, they will appear with a purple outline, clearly distinct from the blue outline (from manually labelled tiles).

After running your first prediction, you can click on the "Config" button to change the number of predictions generated.

Filter by confidence

When you've generated some predictions, our algorithm is more confident about some matches than others, within the number of predictions you desired. As you drag the slider to the right, the confidence filter will increase and the number of tiles displayed will decrease. You'll be shown tiles that our algorithm is more confident are a match as opposed to all 500 predictions, for example. Try dragging the slider to adjust the confidence level!

As you can see, the number of predicted tiles has decreased to 42. Given tiles near the already labeled ones are quite similar, we would expect them to remain, especially when filtering toward higher confidence.

Updating predictions

Once you've added additional positive or negative labels, whether from predicted tiles or other observations, you should update the generated predictions. Given this is a trial-and-error process, being iterative by adding a few more labels and updating predictions is necessary and efficient!

To update, click the "Update" button in the Prediction window.

Watch this video to learn how to run predictions:

How does this work?

Each tile is represented through an embeddings produced by the foundation model. When you select a set of tiles that are similar, we search across all the tiles in this region to find tiles with similar embeddings, or nearest neighbors.

If many points are labeled positive, the machine searches for the average of those positive vectors. If you include negative labels, the machine searches for the average positive vector displaced away from the average negative (by the difference between the average positive and the average negative).

Now that you've learned how to label and make predictions, try building a dataset of some object or feature in the imagery! Go through a trial-and-error process and play around with Earth Index. We're looking to improve the tool, so feel free to share any feedback with us over Discord or at [e-mail].