Scattering Transforms in Python

Please cite:

@ARTICLE{3dfst,
  author={I. {Kavalerov} and W. {Li} and W. {Czaja} and R. {Chellappa}},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={3-D Fourier Scattering Transform and Classification of Hyperspectral Images}, 
  year={2020},
  volume={},
  number={},
  pages={1-16},
  doi={10.1109/TGRS.2020.3040203}}

Based on the arxiv, ieee papers.

Usage

The main file for HSI analysis is hyper_pixelNN.py.

To use this file run the following scripts:

• sh scripts/final/train_dffn_IP_replicate.sh, Replicates the results of DFFN for Indian Pines once.
• sh scripts/final/train_EAP_PaviaU_replicate.sh, Replicates the results of EAP-Area for PaviaU once.
• python scripts/generate_many_dl_runs.py, Generates scripts that will run 10 trials of DFFN and EAP for every dataset.
• scripts/final/IP_fst_svm_all.sh, runs FST and a SVM once for every mask listed in a txt file.
• scripts/final/svm_wst_all.sh, runs WST and SVMs for the 4 datasets mentioned in the paper.
• scripts/final/svm_wst_all.sh, runs SVMs on the raw HSI images for the 4 datasets mentioned in the paper.`

Installation for scripts in scripts/final

• install items in "Software Versioning" below
• compile libsvm in lib/libsvm (run make in lib/libsvm and in lib/libsvm/python)
• get masks
• get mask lists
• get data

Archived version with masks is here.

Downloading Data

You may download a copy of the HSI data mentioned in the paper here

Or, see the GIC website and download the datasets there.

Training Masks

You may download the exact training masks we use in the paper get the masks inside the release here and place them in your masks directory.

When using options like --svm_multi_mask_file_list there should be a txt file that lists the fully qualified path to each mask file that should be used.

Create Custom Training/Testing Splits

Create more training/testing splits with sites_train_val_split.py.

Software Versioning

Tested on Python 2.7.14 (Anaconda), tensorflow 1.10.1, cuda 9.0.176, cudnn-7.0 (8.0 might work too). Red Hat Enterprise Linux Workstation release 7.6 (Maipo). GeForce GTX TITAN X.

conda create -n venvtfnonb python=2.7.14
pip install --ignore-installed --upgrade tensorflow-gpu==1.10.1
pip install sklearn tqdm h5py hdf5storage pillow matplotlib plotly scikit-image

Docker instructions

This docker container was tested on a Nvidia 3090 and a M40.

Setup from scratch

cd ~
git clone https://github.com/ilyakava/pyfst.git
mkdir -p hsi_data/derived

docker pull nvcr.io/nvidia/tensorflow:18.10-py2
docker run --gpus all --shm-size 16G --rm -it -v $HOME/pyfst:/pyfst -v $HOME/hsi_data:/data nvcr.io/nvidia/tensorflow:18.10-py2
pip install sklearn tqdm h5py hdf5storage pillow matplotlib plotly scikit-image
cd /pyfst/lib/libsvm
make
cd python
make
cd /pyfst/masks
wget https://github.com/ilyakava/pyfst/releases/download/MajRev1/pyfst.tar.gz
tar -zxvf pyfst.tar.gz
rm pyfst.tar.gz
mv pyfst/masks/* ./
rm -rf pyfst
cd /pyfst
apt-get update
apt-get install python-tk

Run an example

export DATASET_PATH=/data
export DATA_PATH=/data/derived
cd /pyfst/
CUDA_VISIBLE_DEVICES=0 sh scripts/KSC_fst_svm_one.sh

For this example the ST runs on the M40 in 110s. For the newer 3090, the Ampere architecture requires more graph compilation time because of the older cuda version.

License

MIT