A thread about the newly launched #DynamicWorld landcover dataset by #Google. I had early access and explored this dataset in detail. You may be very excited about this dataset, but likely for the wrong reasons. Sharing some insights, potential use cases, and pitfalls. 1/n
First of all - what is it? It's a Landcover dataset based on Sentinel-2 data - but with a key difference. Rather than a static snapshot, it is a time series. *Every* Sentinel-2 scene is classified with class probabilities for 9 landcover classes. 2/n
https://twitter.com/rebeccatmoore/status/1534982514132996096?s=20&t=Mkc-LFfX-nrr2FaQ7Xj-tQ
It is an incredible technological feat. The dataset contains not just every Sentinel-2 scene from the archive, but every new scene is classified and made available in just a few minutes to all #EarthEngine users a through a dynamic collection 3/n developers.google.com/earth-engine/d…
The classification is derived from a deep-learning model - specifically a FCNN. Rather than a traditional classification product, the output is 9 probability bands for each Sentinel-2 scene - describing probability for each class - giving you per-pixel time-series like below 4/n 
The exciting part of this dataset is its 'dynamic' nature - and availability of class probabilities. It allows you to generate new classification products and analyses easily. It is NOT a ready-to-use classification product. 5/n
If you want a discrete ready-to-use LULC map, Dynamic World falls quite short. Due to the inherent nature of a CNN which includes the context about pixel neighborhood - you lose the ability to resolve pixel-wise differences. Here's a comparison with ESA WorldCover 6/n
Another potential pitfall is the 'Built' class. Contrary to what traditional landcover products may consider built-up, #DynamicWorld class definition is more akin to 'urban' class. That means man-made structures like solar farms in a non-urban areas are not considered built. 7/n
However, the 'built' probability band can be used map urban growth. Since the #DynamicWorld class contains the notion of 'urbanized' pixels -it can be used to map urban sprawl. See my tutorial developers.google.com/earth-engine/t… 8/n
If you are working on crop-classification problems - your work just got a whole lot easier. Capturing crop phenology is as simple as adding the probability bands to the Sentinel-2 bands and training a classifier - without any advanced techniques required for a good output. 9/n
The #DynamicWorld product is amazing at identifying regions that undergo a lot of change. i.e. a few lines of code to identify crops that get flooded regularly or lakes that freeze during winter. Some examples at developers.google.com/earth-engine/t…
If you are excited to explore this dataset - check out this series of tutorials that I developed - with step-by-step instructions and sample scripts to get you started in #EarthEngine developers.google.com/earth-engine/t… 10/n
Lastly, I want to mention that all opinions in this thread are my own and the result of my own exploration. As an educator, I am thrilled that the #DynamicWorld dataset would lower the barrier to entry and enables use cases that required a lot of work before. 11/n
• • •
Missing some Tweet in this thread? You can try to
force a refresh
