NASA Earth Observatory image modified by SkyTruth
Questions about Dynamic World in SkyTruth Alerts
How often is Dynamic World updated?
DW is based on Sentinel-2 satellite imagery. Sentinel-2 collects new imagery at least once every 5 days, but Dynamic World only processes images with less than 35% cloud cover. So in cloudy areas, updates could take weeks or months.
Where does Dynamic World perform best?
According to model developers, DW most accurately classifies relatively clear-cut land cover classes, such as:
- Forests
- Water
- Built area
- Ice and snow
Where does Dynamic World struggle?
At SkyTruth, we’ve also noted:
- Change Detection is greatly affected by snow-and-ice land cover. As the probability of a scene having snow increases, there’s an equal reduction in the probability of other land covers.
- Given the 10m resolution of Sentinel-2, DW can misclassify smaller footprints, especially when in mixed land cover areas. For example, buildings in densely wooded areas will often show up as part of the tree cover.
- DW can have difficulty classifying rocky areas.
Why was Dynamic World added to SkyTruth Alerts?
- While you might be able to apply DW features to a single date, you’ll usually need a date range of a month or several months to get good minimum cloud coverage data.
- Annual date ranges, while slower to process, can smooth out fluctuations by creating average probabilities over a longer period for each land cover classification.
- You can often get a good result by building your own probability filters for Similarity Search or Change Detection. However, inclusion/exclusion of even a single scene (a scene is for a single date) can change those results. Experimenting with date ranges and times of the year is key when looking at local areas.
- Same for different habitats — you might create a nice set of parameters for viewing built areas in the desert, but when you run it in a hardwood forest you get worse results.
- While DW does filter out cloud cover, it is not always accurate. Sometimes missed clouds will present as snow and ice land cover.
- Running a Time Series can be a big hit on resources, so is currently available only at zoom levels 12 or higher. If the time series fails to build, try selecting a smaller date range, a smaller time cadence (e.g., two months instead of six), or a smaller area.
How accurate is change detection and identifying similar features in Alerts?
DW will not perform as well as a dedicated ML model for similarity search or change detection, but it can be very useful in analyzing trends and identifying areas that need further evaluation.
Where can I read more about the Dynamic World model?
You can read a nice introduction to DW in this article by Tanya Birch, Google Earth Outreach’s Senior Program Manager. For a discussion about the methods used by DW modelers to develop and validate the dataset, this is a good read.
Is there a recommended workflow for using your own probability filtering for Similarity Search and Change Detection?
- Start by entering your date range(s) and checking the Show Sentinel-2 mosaic for filtered dates checkbox
- Add a single filter. For example, the Built land cover for a 30% increase in Change Detection. Click Find to see results.
- Check the Show DW pixel stats on map click checkbox.
- Repeat as needed:
- Click a map pixel to see how DW classifies scenes at that location. (For Change Detection, click on the right map while viewing map comparisons.)
- Use this additional information to adjust the existing filter or add a new filter. For example, if pixel stats show a probability for trees of 60% and you want to filter for trees, you should include the Trees land class with a somewhat high minimum probability.
- Click Find to see results.
