Crop non crop Labeling
Labeled images help train and validate the Street2Sat Machine Learning (ML) model that is used to create large datasets. The crowdsourcing process of labeling field-level images as either cropland and non-cropland provides needed data for the AI and ML models to produce accurate crop maps.
Many satellite-based analyses rely on ground-truth or photo-interpreted reference data for calibrating and evaluating models and map-based estimates.
Once generated, an accurate crop mask can be used for a wide range of agriculture and food security analysis and decisions. For example estimating land use change and generating a crop type map.
- Go to collect.earth in your browser (e.g., Google Chrome)
- Click the Login/Register button in the upper right
- Create an account or log in with an existing account
- Request access to Collect Earth Online.
- Navigate to the project link you have been given
Your screen should look similar to the image below.
To make sure you have the correct settings, go through the following steps each time you go to CEO to label points.
- Under Imagery Options in the right panel, select Sentinel-2 - this will be your default image source for labeling. Other sources (Planet Monthly Mosaics and Mapbox) will be used as needed.
- Move the Year slider to the year of your project (e.g., for 2020 Ethiopia Tigray, move it to 2020). This is important since the land cover may change from year to year!
- Move the month slider to a relevant month within the growing season for your project (e.g., for China, move it to something in the range June-October).
- Click the Go to first plot button
When you first go to a point, the window will be zoomed to the bounds of a 1 km2 box around the point to label (yellow box). Your task is to label the point, not the bounding box. You may need to zoom out until you can see the relevant context for the point.
The default visualization selected for Sentinel-2 imagery is “Agriculture”, which is a combination of bands (SWIR, NIR, and blue) to emphasize crops. Growing/healthy crops will usually appear as bright green while fallow fields or bare soil will look brown/orange.
You can select other band combinations to visualize the Sentinel-2 imagery from the Band Combination drop-down. For example, True Color (red, green, blue) is shown here.
Crop calendars specify the months when crops are at key growth stages based on region and crop type. The growing season for most crops in Ethiopia is around June to October.
Planted fields should look green during these months with fields being cleared for planting/from harvesting in the months before/after.
Crop calendars specify the months when crops are at key growth stages based on region and crop type. The growing season for most crops in Ethiopia is around August to October. Planted fields should look green during these months with fields being cleared for planting/from harvesting in the months before/after.
Here are some tips for troubleshooting if you run into issues with these instructions:
- CEO recently added an Update Imagery button that you need to click whenever you change the Sentinel-2 year or month. Make sure you click this after you move the slider.
- Sometimes Sentinel-2 will just show a white screen - this could mean it is taking a while to load or the scene is completely cloudy. Try choosing a different month or panning/zooming to trigger a re-load. You can also try switching to PlanetScope or Mapbox then switching back to Sentinel-2.