Open guy1ziv2 opened 3 years ago
guy1ziv2
commented
3 years ago If anyone want to take see how some potential deep-learning training data we can use, I ingested ~700k field polygons from the France GSAA data of 2019 into Google Earth Engine. The link below provides a visualization over a small area - https://code.earthengine.google.com/57f236e39e1a99d3bcc567741732aea2
guy1ziv2
commented
3 years ago p.s. there are other (more complex/accurate) possibilities for this Hack, but what I would like to test first to get quick results is adapting 'Multi-class prediction with a DNN' example in https://developers.google.com/earth-engine/guides/tf_examples to this problem. I can share a paper I found that used Sentinel-2 to map field boundaries, and a few other example approaches.
craigdsouza
commented
3 years ago Hey @guy1ziv2 , do you and Guido work together? I think I remember you from a hackathon at geo for good 2018 (Dublin). Love your idea, I am definitely not familiar enough with deep learning to take a stab at this yet, but I am starting to follow the literature, field boundaries are super interesting to me, the analogy to OSM is very appealing. If you can point me to good resource papers you know of/ meta papers that summarize approaches on edge/boundary detection that would be super helpful.
Proposal by Guy Ziv -
With the availability of Open Data like OpenStreetmap we want to think basic 'geospatial framework' of the world around us is easily available. In some domains, this is the case - we have decent open data road network, for example. Such data has been used in countless ways, driving innovation and analytics in transport, planning etc. In agriculture, however, we are far off that point - in fact, we don't even have an Open Data of field boundaries and woodland plots. Some European countries, for example, provide some of that data (usually not all land is covered). In others, you can only get that with restrictive data agreements for non-commercial use - usually after months of delay.
In short - we need a way to trace field boundaries and make a free database, updated annually. Planet resolution and temporal frequency can be a great benefit here. First - you can see in Planet features like hedgerows etc. which split fields, but can't be seen even in Sentinel-2. Secondly, in places with lots of clouds having nearly daily images make it much more likely to get ANY image.
What would a hack team for this project work towards in 2 days?
Experiment in a region of the world where we already have data (I got field-polygons from a number of European countries that can be shared). I would suggest to test a Machine Learning pipeline of some sort - but details should be discussed as a group
How would this project use Planet's data & platform?
Firstly, it will use Planet imagery... obviously. I think an interesting product are ready-make composites (monthly ones?). I know it is possible to get Planet download directly to Google Cloud Storage, from which you can ingest into Google Earth Engine, and you can run a trained CNN in GEE now. Never tried any of those steps....
What obstacles or blockers do you think this project might run into?
Access to the specific APIs - don't know what we can play with. Also not sure how to do the Machine Learning part - either in something like PyTorch or within GEE? It may end up too resource intensive as well.
And finally... pitch your idea: why should people hack on THIS project in particular?
The first reaction many people (from other fields) have is "really ? there is no free data of fields ? this is so easy to see in Google/Bing/XXX maps !" So no - there isn't. But there can be!
Secondly, there is great scientific value in such a map, and if we can trace the change in field size over time it also give us indicator of structural changes (sorry for the jargon). Having those polygons opens up a range of other things from object-based classification of crop types, to verification of farmers reporting etc.
Finally, I even think there is commercial value in such a database, Similar to the value of a road network geospatial framework - given that all sort of precision agriculture and farmer tools can use such a database, I can imagine companies would pay for it if that existed and is good!
If you're interested in hacking on this project, add a 👍 reaction to this post.