Comment by notahacker | Hacker News Reader

Comment by notahacker | Hacker News Reader

Parent Full thread

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.

Thanks! This analysis is really helpful to sense check the accuracy of these kinds of ML predictions based on satellite data.

Out of interest, could you share any further insights around the challenges using the Sentinel-2 data set?

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.

Reply on news.ycombinator.com