CGIAR Crop Yield Prediction Challenge
$3,000 USD
Can you predict maize yields on East African farms using satellite data?
514 data scientists enrolled, 145 on the leaderboard
AgriculturePredictionComputer VisionUnstructuredImageSatelliteSDG2
Kenya
21 October 2020—7 February 2021
Ends in 21 days
Any CNN based models
published 7 Jan 2021, 08:12

Did anyone try out building a CNN regression model using the 360 channels? If so was there any good RMSE score. For me the RMSE looks very odd

Try a lstm over a cnn. That is 30 channels spread over 12 months. So lstm should hopefully capture time

Thanks for the suggestion @aninda_bitm I will give it a try

Yes.... The score is very odd like you said. 1.7### on LB.

I would think. That's not an odd score. May need

Really........ Actually with DNN i got 1.7300 on LB without any hyperparameter tunning and any other adjustment. The score with CNN is not that good at all compare to DNN.

1.7 far better than what I am getting around 20

Really........ Actually with DNN i got 1.7300 on LB without any hyperparameter tunning and any other adjustment. The score with CNN is not that good at all compare to DNN.

The CNN architecture, optimizer, and learning rates had to be changed to finally end at 1.8 on LB. So there is room for improvement using CNN.

BTW anyone tried out removing the low-quality images, if so what was the best LB score you have? for me, it was around 1.81 based on the starter notebook.

Great...... Like i said before my CNN didn't give me good score while my DNN did very well (1.7300 on LB) even though am yet to change/adjust my optimizer and learning rate respectively if that would further improve the score. For me removing low-quality images worsen my score on LB. I mean removing "low-quality images" have no positive effect on my score at all.

@saibhargav08 I am just a beginner and am learning. Please could you show me how this 360 band is being fed into a CNN. I haven't even made any submission only how the bands are being put into a CNN thanks. Just that