SkyTruth Showcases Novel Use for Dynamic World Dataset at Geo for Good
A new, free collection of data—in tandem with Alerts—bolsters our ability to detect habitat change in near real time
This Thursday (October 6) SkyTruth’s Chief Technology Officer Jason Schatz will speak at Google’s Geo for Good plenary session alongside conference organizers from Google and World Resources Institute (WRI). Jason will present technology developed at SkyTruth that focuses on the value of Google and WRI’s new Dynamic World (DW) dataset. DW is a near real-time, land use/land cover (LULC) dataset that we’ve incorporated into SkyTruth’s Alerts application.
The unique data—freely available to the public—make it possible to create algorithms that automatically detect specific habitat changes, such as new buildings, deforestation, desertification, and changes in surface water boundaries. When coupled with existing features of Alerts, the combination can yield huge benefits for land use monitoring. Those interested in viewing the plenary session can sign up to attend Geo for Good virtually.
Dynamic World (DW) is an exciting product in many ways.
- It’s produced from a machine learning model that processes images from the European Space Agency’s Copernicus Sentinel-2 mission. Depending on location, you may see new results every 2–5 days, hence the name “near real time.” Traditional global land cover datasets, on the other hand, are produced annually at best.
- It has a 10-meter resolution, which means land is classified for every 10-by-10-meter (10x10m) block on the ground. Many of the prior land classification datasets were based on Landsat imagery, which had a 30-meter resolution, or one land class for each 30-by-30-meter (30x30m) area. Ten meters has nine times as many pixels as 30 meters, which means nine-times-better detail while viewing the data.
- Most land classification datasets will assign a LULC value to each pixel. DW assigns a probability to each of its nine LULC classes for each pixel. So we know the probability that the pixel for a given date represents a building, forest, water, or any of the other DW classes. This is important data when writing algorithms that tease out the contents from an image.
Why is all this important? The short answer: It’s a lot more data. It’s also the right kind of data, and by making DW an Earth Engine dataset, the data is easily accessible to developers.
An automated change detection algorithm is one of the most useful tools that we can provide to environmental groups, government agencies, and individuals who monitor land use. One approach to creating change detection requires labeling scenes (satellite imagery taken on a specific date) and seeing where those labels differ from one date—or, more likely, a range of dates—to another. Deep learning has been used successfully to create these labeled scenes, but up to now most of the available datasets have been for annual composites of images, where a pixel’s LULC label represents a single classification for the entire year. In such cases, much of what occurs on a property throughout the year gets lost. For example, a single 10x10m area might transition between cropland, grassland, bare, and buildings all in the same year. An annual LULC label can only assign one label to that pixel, thus missing the full picture
DW solves this problem by applying LULC labels to each scene. It then goes one better by assigning probabilities to each LULC category for each pixel. Probabilities are important because of the imprecise nature of computer vision in general; identifying crops, bare ground, grassland, or shrubs can be tricky work for a computer.
All things eventually change, so a perfect change detection algorithm would detect change everywhere. The hard part is defining what changes matter, over what time period, and measuring those changes effectively. If I were to build an algorithm that perfectly detects broad vegetation change, that wouldn’t help me locate a desert. In the same vein, if I only care about forests—and not grasslands or crops—then building a forest change-detection algorithm would overlook shrublands.
Keeping all this in mind, there are a few important questions to ask yourself before you can build a useful change detection algorithm: 1) What changes do different stakeholders care about? 2) What ecosystem are we working in and what do those changes look like there? 3) How do we measure those things reliably and distinguish them from changes we don’t care about?
Using SkyTruth Alerts, we address the need for habitat-specific change detection by going one step further: making it specific to a user’s geographic area of interest (AOI).
Here’s a workflow for creating AOI-specific change detection using DW in Alerts.
- Go to an area where you already know changes have occurred. This is your playground for testing various DW bands and probability thresholds.
- Select the Explore change detection parameters layer for exploring DW, then start by entering before-and-after date ranges. Again, the idea is to use date ranges that match when you already know there have been changes. That will help you play with the DW settings to match those known changes.
- Check the Compare using Planet basemaps checkbox. This will bring up basemaps with before-and-after, high-resolution satellite images in a comparison slider to help you evaluate results.
- Start with All Land Classifications at 25%, then Run the layer.
- Use the Pixel Stats layer to examine false positives and false negatives. Are there any bands that obviously increase from before to after scenes? Are there minimum values for certain bands in the after scenes? Or maybe there is a band that fluctuates widely and you want to filter that out?
- Adjust the Explore change detection parameters settings based on what you discover in Pixel Stats, then re-run the layer.
- Repeat steps 5 and 6 until you see satisfactory results.
- Save those results and run again in the future as needed.
Using this workflow to experiment with DW data, SkyTruth identified DW parameters and values that compare two date ranges, allowing us to detect these changes:
- New buildings developed in a northeast suburb (found by looking for an increase in the BUILT land cover of at least 25%).
- New oil drilling in the Permian basin (found by looking for an increase in the BARE land cover of at least 20% AND having a BUILT land cover of at least 5%).
- New unauthorized mining detected in the Amazon forest (found by looking for an increase in the BARE land cover of at least 25% AND ignoring large fluctuations in the WATER land cover).
These detection algorithms aren’t perfect, but they don’t have to be; they just have to point to where you need to look closer. With its built-in satellite viewing and comparison tools, SkyTruth Alerts is designed to help you take that close look.
Note that this approach will always be outperformed by a bespoke model designed to detect a specific type of change in a specific habitat. For example, the new mining identified in the Amazon is not as accurate as machine learning models that are built specifically for this task, such as the one we built at SkyTruth. But DW can still be a useful addition because it’s so quick. At SkyTruth we’re using DW data to improve our Amazon mining model by removing some false-positive detections; for example, we’ve filtered out agricultural fields after harvest because they have similar visual signatures as mines.
Roadmap for DW in Alerts
Everything we’ve done to date is still being tested, and we welcome partners! If you’d like to join the experiment, you’ll find some guidance here.
As we continue to work with the DW dataset, we are considering some of the following enhancements to Alerts:
- Making it easier on users by creating a dedicated form or wizard to walk users through exploring DW data
- Generating Monthly Change Detection alerts—sent straight to your email inbox—for models that produce good results
- Auto-generating images for changes that were detected using a variety of bands and thresholds, then allowing the user to decide which one works best going forward
As always, we want to hear from you! Your feedback is an important consideration in deciding which tools we develop at SkyTruth. Contact us at firstname.lastname@example.org with ideas, use cases, and questions.