ForestLoop is an architecture to train supervised classification models for land use land cover that are locally calibrated for a project area. The architecture enables a GIS expert to tag satellite imagery, run the classification, and iterate on that process until the classification is adequate.

1. Load & Configure Forest Loop

To load ForestLoop open the Analysis Modal and in the Land Cover section select "Land Cover (single year)".

This expands the computation options to configure including the start date, end date, data source, sample radius, and satellite composites. The configuration affects the timing of imagery selected, how the model chooses tagged points to include in training, and which data the model is trained from.

The "Data Source" chooses which labeled points to include:

• Auto: automatic algorithm without user tagged points
• Polygon: your labeled points inside the polygon + sample radius
• Crowd-sourced: all user-labeled points inside the polygon + sample radius
• Global: use all verified global points

The "Sample Radius (km)" changes the search radius for fetching sample points from the database around the geometry centroid. This is applicable for "Data Source" types "Polygon" and "Crowd-sourced".

The Satellite Image Composites chooses which land composite map to run the algorithm on. Options include:

• Landsat 8
• Sentinel-2
• Alpha Earth Embeddings

After making selections to configure the computation, click the "Run analyses" button to start the computation.

The computation will move through Queued -> Pending -> Completed. Once the computation is "Completed" then it is ready for use.

Click on the computation to enter ForestLoop.

Upon entering ForestLoop, the left Summary panel will read "No widgets found" and the main map will display a classification. This classification is the default ForestLoop classification model. In some cases, this will be a sufficient land use land cover model. To start labeling points, select "ForestLoop" in the left panel.

Toggle between the LULC classification and the high resoluiton basemap using the eye icon in the Map Settings, Active Layers, "Land Cover Map". This map will also contain the year that the computation was configured for.

2. Prepare Satellite Data

To label points, satellite imagery from a defined source and time period must be loaded. The basemap provides a high resolution reference but may not be from a time that is relevant to the classification.

To add the satellite imagery to the Active Layers, navigate to the Available Layers section and select the relevant imagery and time window.

3. Label Points

To start labeling points, toggle the satellite imagery to be used as the composite using the eye icon. In the map panel, click Edit Points. This opens the labeling toolbar. The labeling toolbar contains 10 LULC classes:

• Degraded
• Forest
• Water
• Built
• Savannah
• Clouds
• Degraded Forest
• Bare
• Grassland
• Shrubland

The label sizes can also be customized with the second dropdown selector. Choose an LULC class and start labeling points. The user can shift between satellite composites and other datasets to assist with labeling.

After labeling points, click save points. The number of points labeled in each class will display in the left panel. To remove points, click back over a previously labeled point.

4. Retrain Model

When labeling is complete, retrain the model using the labeled data. This can be found at the bottom of the left panel "Update map classification". While the map classification updates, the map layer for "Land Cover Map" will not be accessible. After the classification updates, a new "Land Cover Map" will appear in the Active Layers.

Iterate through the label and retraining steps until the LULC classification is adequate. Tagging points required specialized GIS expertise to recognize classes. Earthshot can provide customized pricing for our in house GIS team to label training points.

5. Access Results

Results export is done through the report builder page. To access the report builder, go back to the polygon view and select "Report". This load the empty Report Builder. Use the template for "Land Cover (single year)" in the templates dropdown or drag the widgets from the Available Widgets panel. Now the results are displayed in the Report Builder.

Click "Save report" at the top and then open the dropdown beside "Download PDF report". Select "Land Cover Map" from the Export viewport data.

Now the data gets prepared for export. This step can take several minutes.

The Exports drawer on the right side of the screen tracks exports. Initially, this will have a blue dot for Processing. When the data is ready to download, the color will change to green and the geoTIFF is ready for download.

Tips

Tips to improve results:

• Define LULC classes by overviewing the high resolution basemap and deciding what is important to map.
• Overlay the canopy height models to detect certain forest presence.
• Overlay the Sentinel-2 RGB and infrared false color layers. Try to detect the same LULC structures present in the high resolution basemap in the coarser resolution image.
• When placing the samples it is important to pick up pixel groups that are as homogeneus a possible. Meaning its better to sample the core of a cover rather than its limits where it could be blending into another class.
• It is also good to try to keep balance between classes in number and size of samples.