Crop Yield Prediction USA results

[gabrieltseng]

Hi! Thanks for this collection of datasets and benchmark models.

The Crop Yield Prediction model has very different yield values for the US and Brazil.

For example, the test yield values in Brazil have a mean value of 2.69375, while the test yield values in the US have a mean value of 44.39524599226092.

I have two questions:

• Is this expected?
• Are the results reported in the US for the You et al. model for these yield values? I specifically ask because the RMSE of 0.37 is much lower than what is reported in the referenced paper (e.g. the CNN+GP has an average RMSE of 5.55). If what is on the leaderboard is correct, what changed to improve the results so much?

Thanks again!

Gabi

[gabrieltseng]

^ I've figured out that the USA data is in bushels per acre, while the leaderboard reports tons per hectare. Using the conversion 1 bushel/acre = 67.25 (67) kilograms/hectare, I get the results on the leaderboard, since

(5.55 * 67.25) / 1000 = 0.37

I have a follow up question:

• The 5.55 result seems to be taken directly from the original paper, which reported results for 2011, 2012, 2013, 2014, 2015. However, the test years in the CropYieldDataset for the USA contain the following test years:

array([2016, 2007, 2005, 2008, 2013, 2010, 2009, 2011, 2012, 2014, 2015, 2006])

Was the model rerun for all the test years, or are the results in the leaderboard only for a subset of the years?