Indonesia Deforestation Segmentation

Top 1 solution for Indonesia Deforestation Segmentation Challenge. (0.379 Dice Score on LB)

Additional Data link

In this work, I use a public dataset and add it into the provided dataset. You can download the data here: ForestNet dataset
The additional data contain the mask as forest_loss_region.pkl files, so I processed to png files in eda2.py. After running this python code, you will get the png mask-image files as well as the csv file containing the annontation of them.

Pretrained Weights

You can find the pretrained model here: link

Model Used in the Challenge

I used UNetPlusPlus with NFNet backbone, and I used visible and infrared images as the inputs.

How to Run

For training:

python training.py

For validation:

python inference.py

Technique

[x] UNetPlusPlus (and other models you can find in my code)
[x] TTA (geometric and pixel level TTAs)
[x] Ensemble (In the scope of this competition, I did not use ensemble but I have included the ensemble code in my code)
[x] Models using auxilary information (latitude, longtitude, year)
[x] Use SegFormer to finetune
[x] Cutmix
[x] Mixup without hesitation
[x] KFold training

Further Improvement

[] Heavy TTA. (zoom-in, rotation more)
[x] Postprocess (Morphology, Connected Components)
[] Use SAM to finetune the model

Reference

Challenge website
ForestNet

Author

Thanh-Tin Nguyen