Skip to content

Latest commit

 

History

History
51 lines (25 loc) · 2.28 KB

README.md

File metadata and controls

51 lines (25 loc) · 2.28 KB

STAC-Overflow

1st place solution for STAC Overflow: Map Floodwater from Radar Imagery hosted by Microsoft AI for Earth https://www.drivendata.org/competitions/81/detect-flood-water/

Disclaimer

Many thanks to the organizers from Driven Data and Microsoft AI for Earth for such an interesting competition and to the participants for the intense struggle until the last minutes!

If this solution seemed useful to you, be sure to share ⭐️

https://t.me/hahatons (the channel is in Russian, but we can discuss something in English)

About the idea of the solution

Initially, I understood that I would not be able to build a super complex neural network, because either there was not enough knowledge or there was not enough computing power.

Therefore, the only chance to win was to come up with a simpler method for determining flooding. To do this, I studied articles about how waterlogging is determined now. There were neural network methods, but there were also mathematical methods. From which I concluded that in addition to segmentation by a neural network, you can try to determine the flooding pixel by pixel by some formula.

But since I am a "cool" data scientist 🦧, I did not output the formula manually, but trained ML models – Catboostclassifier, which solved the binary classification problem on pixel-by-pixel data.

Before that, I also trained the Unet models.

Further, I noticed that the models often do not fill the necessary zones, rather than overfill. Therefore, I combined the predictions of these two approaches, taking their maxima, not the average.

And as you can see, this approach worked and brought me such an important victory! 🥳

You can see other notes about the solution in the jupyter-notebooks.

Solution plan

  1. load_external_data.ipynb

This notebook is downloading additional data from Planetary Computer. Spoiler: Nasadem band is an incredibly important

  1. catboost_model.ipynb

This shows the preparation of pixel-by-pixel data and the training of CatBoostClassifier models on them.

  1. unet_model.ipynb

Here is a classic segmentation approach using neural networks with the Unet architecture with EfficientNet backbone.

  1. compare_methods.ipynb

This notebook shows a comparison of the results of the two approaches and their combination.