Sar UNet

This repo has a simple training and evaluation script for using a UNet with SAR data.

How to Use

• Images shold be placed in data/imgs/ and ground truth should be placed in data/masks/.

  • Note that the name for an image and its associated mask should be the exact same. E.g. data/imgs/img1.npy and data/masks/img1.npy

• data/txt_files/train.txt and data/txt_files/test.txt should contain the names of the traing and testing images respectively, with each image name on a new line. Example below:

  img1.npy
  img2.npy
  img3.npy
  

• Install requirements: pip install -r requirements.txt.

• Install package lakeseg package: pip install -e ..

• Run training script: python lakeseg/train.py -d path/to/data/ -ch 2 -cl 2.

  • Note that the above command assumes two ice classes (ice and water) and two SAR channels (HH and HV).
  • Training info can be found at the wandb link produced by the script.
  • Checkpoint are saved in temporary lakeseg/checkpoints/ directory.

Generating Data

• The scripts/make_mask.py script creates UNet compatible masks from ice_chat.dat files.
  • ice_chart.dat files contain lat/lon points with an associated sea ice concentration.
  • This script creates a contour mask from the lat/lon points, converts the contour map into a geotiff, adds projection info to a SAR geotiff with ground control points (GCPs), clips both geotiffs using a lake shapefile (found in data/shapefiles/), and saved both the SAR image and mask as a .npy files.