An experiment to test the use of the pre-trained Prithvi model for a change detection application
The test case is the xBD dataset containing before and after imagery of building damage caused by natural disasters.
The idea is to use the the pre-trainied Prithvi encoder to embed the "pre" and "post" image sequence and train a semantic change detection head that takes the embedding as input.
We can compare performance to the AB2CD results on the same data.
pip install -r requirements.txt
Get the model code:
git clone https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M
mv Prithvi-100M prithvi_100m
Download the checkpoint file, Prithvi_100M.pt, manually from: https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M/tree/main
Make it importable:
touch prithvi_100m/__init__.py
Download the data set and follow the instructions at: https://xview2.org/download
python create_dataset.py target-masks --data_dir path/to/data
Sentinal bands in HLS data used to train Prithvi:
B02: blue, 490 nm
B03: green, 560 nm
B04: red, 665 nm
B8A: IR, 865 nm
B11: SWIR, 1610 nm
B12: SWIR, 2190 nm
Loss Scaling
- Over the dataset, compute the pixel area of each class for each image. (Done in create_dataset.metadata and stored in metadata.csv)
- Sum the pixel area of each class over the whole dataset.
- Find total pixel area of the dataset,
total_area = num_images * 1024*1024 - Calculate the loss weights as:
inverse_probability = 1 / (class_areas / total_area)
weights = inverse_probability / (inverse_probability.sum())