Appreciate it! Great to hear that the notebook is of help to you.

Convnext base was clearly the outstanding backbone per my small experiments. I did'nt explore >3 channels - would be something worth exploring to see how it stands out. I don't understand your band order - could you kindly explain how you arrived at that?

Lastly, did you attempt object detection using geometry information? if yes , how were your results?

So, for each sample, it contains 12 TIFFs (one per month), and each TIFF has 12 spectral bands—so for a given sample, we have 144 grayscale images total. I treated each unique (month, band) pair as a separate input channel, resulting in a 144-channel input tensor per sample. This structure allows the model to learn both temporal and spectral patterns jointly.

For object detection—I didn’t explore it. The spatial resolution of the images is quite low, which makes precise localization of objects challenging. So I focused on approaches that leverage dense classification instead.

Thank you.

For the boosting model (specifically LightGBM), I didn't need to fill the NaNs since it can handle missing values natively.

But for the ConvNeXt-fused model, I filled all NaNs with 999999 to explicitly signal that the data point is out-of-distribution. For missing months, I used blank (zero-filled) images so that each sample consistently had data for all 12 months.

Thanks @NogginHq

The metadata and geometry features were only used for EDA and visualization purposes, not for model training or inference. I've updated the notebook to clarify this more explicitly.

Only statistical aggregations and vegetation indices were used

Awesome. Thanks for the clarification.

Thank you!

Thank you! Pseudo labelling gave me quite a boost, CV from .96 to .98 and private lb from .95 to .97

That is huge!

Thanks! Congrats to you and Muhamed_Tuo as well.

I initially used 10 stratified folds, but the CV–LB correlation wasn’t reliable. I switched to 15 stratified folds based on the assumption that the public LB was around a 25–30% split. With only 282 test samples, 15 folds gave me a closer mimic of that split and improved consistency across CV and LB.

Thank you! Hope you find it helpful.