Farm Pin Crop Detection Challenge
$11,000 USD
Classify fields in South Africa by crop type using Sentinel-2 satellite imagery
4 March–15 September 2019 23:59
796 data scientists enrolled, 42 on the leaderboard
Many fields defined outside of satellite image
published 22 Mar 2019, 15:43
edited 1 minute later

While investigating the data (particularly the file), I noticed that nearly 10% of the fields in the test and train data shapefile do not overlap with the images provided in (I suspect the same problem is present with the other raster image files). I have tested in QGIS and a few other programs and come to the same result -- near the South-East of the data a portion of the test and train fields continue beyond the bounds of the satellite image. Has anyone else encountered this problem? Or am I just very confused?

Example field id's affected are 1, 3, 5, 6, 3399, 3246

I have the same issue, anyone have an explanation ?

Anyone have advice for handling these cases?

Ditto! I find training data index values 0:13, 47, 67:84, 158:202, 1482:1505, etc. do not overlap the raster files. Is this possibly an error in re-projecting the shapefile into the same CRS as the images?

I too have encountered this issue. For now, I am simply assigning the same probability value per crop type so I can pass the submission step for these 'error' rows, but will have to find another strategy. I find it hard to believe some of the scores on the leaderboard if everyone is meant to only use the provided data. 8% of the test polygons fall outside of the raster extents.