Checking this out in a few areas near me:
1. Many golf courses are identified as maize or wheat fields.
2. Nature preservation areas are generally identified as grass or maize.
This "AI-based approach" seems to be off by at least an order of magnitude. A great idea but it certainly needs serious refinement before use by any government or company.
I think they probably need more regional adjustments for the probability calcs I assume underpin their output. The UK Department for Environment, Agriculture and Rural Affairs, for example, has good open data statistics on crop growth, and stats are likely to be reasonably accurate since data is reported for subsidy purposes, and farmers are well aware agricultural fraud monitoring takes place.
OneSoil suggests 325k hectares of UK land dedicated to potato growing in 2017, over twice as much as the 145k reported by DEFRA; 378k ha of maize cultivation vs 195k according to DEFRA.
Other figures from OpenSoil such as barley and wheat are much closer to the reported figure and so may be fairly accurately categorised by its ML process, and some figures are going to be hard to fairly compare due to different categories or multiple crops per year)
At the same, having had a former colleague work on using Sentinel-2 data to classify land use, I'm well aware it's not an easy problem to solve.
Happy to share my thoughts on open data sources and validation steps in more detail by email if you like if you're working on this or something similar.
Not sure my observations on the Sentinel-2 data specifically are going to tell you anything you don't already know though - my original comment basically meant that identifying heterogenous and changing land use against heterogenous and changing surrounds on a regional scale is particularly hard when your spatial resolution is low enough for land areas being categorised to often only be a handful of pixels, and harder still when potential calibration metadata is older than the earliest images. We (my colleagues rather than me) solved our problem by incorporating other lower res and even radar datasets at the initial identification step and then having a manual verification stage using high res optical to evaluate the model output, but we were only identifying a few very specific rare types of land use. And again you're probably aware a model that's well fitted to one region might well needs recalibrating when used on regions with different topography and climate and typical land use patterns.
Checked Hungry, where it stated soybeans is the top third crop, but in reality, soybeans are marginal there. To find outliers like that it would be interesting to compare data with published statistics.
It's identifying 9.4 ha of grapes in my neighborhood in what I know is a public park. Also identifying alfalfa in areas that are simply unforested California open space.
Works surprisingly well for "rural" Southern Germany, Black Forest area where I grew up. It's hilly, so most of the area is forest or grass - but the few fields I know are recognized and mostly with the correct crops as well.
I have a large team that has been working on computer vision driven solutions for 6 years now. Getting something like this working is easy* now (just 6 years ago, not so much). But getting it to have any reasonable accuracy rate is beyond hard.
*easy is probably too strong a word. It is straight forward.
It looks like the classifier has a lot of trouble with water bodies. Looking around through the Puget Sound area (near Seattle) there's a whole lot of fields that would be underwater.
Thanks for checking it out! We indeed had a problem that water bodies were recognized as some specific crops and we're working on fixing it in the new version of the algorithm :) If you spot any more locations, please send them to hello@onesoil.ai, we really appreciate the help!
Interesting project. It really should be compared against some databases of what crops are actually being grown to get a better sense of what one might take away from it. Just spot checking local cases around the Bay area seem to be about 40 - 50% accurate. But there are also issues where crops are rotated so I can imagine one issue would be getting crop images from different seasons could throw it off as well.
That it presents numbers as if it were 100% accurate, is an issue for me.
Thank you! We compared the results of our algorithm to the datasets we were able to get and the accuracy of crop classification, or F1 score, is 0.91. This is stated in the "How it works" section of the map: https://map.onesoil.ai/2018?about. Hope this answer is helpful :)
Around LA the map is off by a lot. "Nuts" in the National forest wilderness, "wheat" in residential neighborhood. But I think it shouldn't be too hard to fine tune this with some level of manual labor. I bet a lot of the models need regional customization. You can't compare dry areas in the West of the US with ares like Austria.
It's a beautiful map but wildly inaccurate as others have pointed out. I compared some of the results against plots of land that I'm familiar with and found it to be off by orders of magnitude in all cases even for areas that are known for certain cash crops (e.g, Assuming every vineyard in Napa Valley is Alfalfa instead of grapes). Meanwhile, my family's small vineyard in the central valley is categorized as as soybeans and the orchard next door (one of the only crops that's clearly identifiable via satellite) is identified as wheat.
Given the prevalence of "grass" in national forests, I suspect it is the kind that has only been legalized at a state level. If the folks growing the "grass" find out they can be detected, especially before their "crop" is ready for harvest, they might be unhappy about it.
I am from the OneSoil team. Thanks for your comments! We appreciate your feedback a lot. Currently, we're working over the algorithm improvements and it would be great if you can share the exact coordinates where the crop is detected incorrectly. This will help us a lot.
We're also interested in cooperation with those specialists who can provide ground truth data about the field borders and crops growing on them. So. if you know someone, please share this info with us (our email is hello@onesoil.ai)
19 comments
[ 3.1 ms ] story [ 46.4 ms ] threadOneSoil suggests 325k hectares of UK land dedicated to potato growing in 2017, over twice as much as the 145k reported by DEFRA; 378k ha of maize cultivation vs 195k according to DEFRA.
Other figures from OpenSoil such as barley and wheat are much closer to the reported figure and so may be fairly accurately categorised by its ML process, and some figures are going to be hard to fairly compare due to different categories or multiple crops per year)
At the same, having had a former colleague work on using Sentinel-2 data to classify land use, I'm well aware it's not an easy problem to solve.
Out of interest, could you share any further insights around the challenges using the Sentinel-2 data set?
Not sure my observations on the Sentinel-2 data specifically are going to tell you anything you don't already know though - my original comment basically meant that identifying heterogenous and changing land use against heterogenous and changing surrounds on a regional scale is particularly hard when your spatial resolution is low enough for land areas being categorised to often only be a handful of pixels, and harder still when potential calibration metadata is older than the earliest images. We (my colleagues rather than me) solved our problem by incorporating other lower res and even radar datasets at the initial identification step and then having a manual verification stage using high res optical to evaluate the model output, but we were only identifying a few very specific rare types of land use. And again you're probably aware a model that's well fitted to one region might well needs recalibrating when used on regions with different topography and climate and typical land use patterns.
Works surprisingly well for "rural" Southern Germany, Black Forest area where I grew up. It's hilly, so most of the area is forest or grass - but the few fields I know are recognized and mostly with the correct crops as well.
*easy is probably too strong a word. It is straight forward.
That it presents numbers as if it were 100% accurate, is an issue for me.
I am from the OneSoil team. Thanks for your comments! We appreciate your feedback a lot. Currently, we're working over the algorithm improvements and it would be great if you can share the exact coordinates where the crop is detected incorrectly. This will help us a lot.
We're also interested in cooperation with those specialists who can provide ground truth data about the field borders and crops growing on them. So. if you know someone, please share this info with us (our email is hello@onesoil.ai)