Primary competition visual

Côte d’Ivoire Byte-Sized Agriculture Challenge

Helping Côte d'Ivoire
7 000 EUR
Completed (10 months ago)
Classification
Object Detection
894 joined
163 active
Starti
Apr 25, 25
Closei
Jun 16, 25
Reveali
Jun 16, 25
User avatar
Koleshjr
Multimedia university of kenya
Private LB 2nd Solution
Notebooks · 17 Jul 2025, 07:59 · 10

Too lazy to write a writeup, hope you're good at reading code! If you have questions, I will gladly answer them :)

/catt_byte_sized_challenge_image_satelite_modelling-2.ipynb

Discussion 10 answers
User avatar
CodeJoe

@Koleshjr Thanks for sharing. I have a few questions to ask😅.

1. How much boost did the robust scaler add to your score? I normally train on the raw values when it comes to gbdt.

2. On my side, some indices turned to worsen my results, only GNDVI, NDVI and CI did its best. Did I do something wrong?

3. What different things did you do to get a very high private score (almost touching 0.98, that's amazing bro🔥🔥)?

17 Jul 2025, 09:18
Upvotes 0
User avatar
Koleshjr
Multimedia university of kenya

1. The improvement from robust scaler had no significant improvement , it was negligible

2. If you look at the feature importance plot , we only focused on median alone. So FE with median features led to significant improvement

3. Tricks to get high score: Patching plus median features

User avatar
CodeJoe

Oh alright makes sense,

Could you elaborate more on patching?

User avatar
Koleshjr
Multimedia university of kenya

You don't use the whole image to create features , you create a crop of it of size NxN where N is Tunable

User avatar
CodeJoe

Ohhh okay! This is definitely new to me. So you focus more on the center of the image or anywhere the phenomenon is mostly located.

Thank you for taking time to answer my questions. I am really grateful and keep winning!

User avatar
Koleshjr
Multimedia university of kenya

I have talked about it in previous competitions:

2nd Place Solution - Zindi

Thanks

User avatar
CodeJoe

Thank you @Koleshjr. Learning a lot from you!

User avatar
Koleshjr
Multimedia university of kenya

welcome, and thanks for your questions I think this discussion acts as the writeup I was too lazy to write😅. Congratulations again for your recent win and I wish you more wins to come!

User avatar
CodeJoe

Thank you big man😊! More wins for all of us😅.

@Koleshjr Thanks for sharing this. Can I please get the code that generated these 2 datasets "train = pd.read_csv(new_path + "train_monthly_stats_patch_8.csv")

test = pd.read_csv(new_path + "test_monthly_stats_patch_8.csv")". I will appreciate it. I just want to learn what approach you took.

25 Jul 2025, 13:59
Upvotes 0