# segmentation approach

I treat this problem as a segmentation task with MSE loss.  My base model is SwinUnetr from monai framework. A few data augmentation technics are used: random flip, gaussian noise, grid distortion.
An input is resized to 224, but the loss is calculated on original size image after resizing. The solutions compose of 2 stages.
The first one uses wind speed (ustar), temperature (theta), and fire-line location (xi), slope (alpha) and mean wind speed (u10) of first 5 time steps to predict 20 next line location (17 input channels, 20 output channels).
The second stage uses previous 25 fire-line location, alpha, u10 to predict the next 25 fire-line location (27 input channels, 25 output channels).
An output of the previous prediction is in range 0-1. In order to binarize and enhance the signal, otsu threshold per frame is applied. I apply post-processing with use of otsu threshold and a binary dilation to zero background (places where is no fire). The final output are floats in range 0 - 1.
[Notebook](https://www.kaggle.com/code/jankowalski2000/segmentation)
There are still many options to check:
dice loss, tversky loss, combo loss with MSE
more aggressive augmentation methods
learning rate schedulers
I hope this post makes me discussion expert :)