Hello Fellow Zindians,

Sorry this took too long. Thank you @Zindi and INEGI for such an amazing competition. My solution was purely gbdt models. All gbdt models i.e catboost, lightgbm and xgboost could achieve second place without any ensembling therefore I never ensembled and went only with Lightgbm. I tried deep learning models but they took a long time to train and I wasn't patient enough.

Key insights

• Balancing the distributions equally by undersampling
• Patching helped prevent overfitting (something most people here faced including me when I first started). That is why you were getting high ROC scores and does not translate well in the public leaderboard
• Vegetation Indices for the win

Here is the Link to the github repo:

koleshjr/INEGI-UN-GGIM-Human-Settlement-Detection-Challenge-by-ITU: Can you create a satellite-based detection solution for human settlements?

Please leave a star if you find it interesting as it motivates me to open source more of my work.

Thank you!