Dirty data maybe?

So so right 🤣

• Test and Train have diffrent distributions .
• Small data ( model can't learn easily ), in last year competition (train has 87K rows ) I was able to develop two neural networks that are fairly similar to gbt models. This year's performance is abysmal.
• Also Time is important as mentioned here :

Looking at these images, it was evident that the crop got harvested sometime in March or in April starting, which is expected as the wheat crop duration is 100 days, and in most of India, it gets harvested after March. So all the images after harvesting basically capture empty fields, which may be why the composite is not showing vegetation. This means we're basically classifying empty fields, as we lost the important information about the crop by making a composite with the empty fields.

great insights. Also @curiousmonkey7 That was an amazing discussion btw

Thanks.

"Time is important", @ASSAZZIN ,have you taken time into account while training? because I think it's the same for all images.

No, if I'm not mistaken, All observations has the same date!

Yeah.

The pain of seeing your submissions getting worse and worse when you use what's supposed to be the right cv for such problems 😂😂😂

😂😂😂

How are you guys even getting less than 1.15 😂. Anyways let's wait for this competition to end so that we may know what you guys used😂. 1.11 is crazy amazing

😂😂😂

😆😆

🤣🤣🤣🤣