Awesome

Improved Difference Images for Change Detection Classifiers in SAR Imagery Using Deep Learning

This repository includes code for the paper Improved Difference Images for Change Detection Classifiers in SAR Imagery Using Deep Learning. Published in IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 61, 2023, https://doi.org/10.1109/TGRS.2023.3324994 (open-access, also available in the paper folder in this repository).

Download the dataset

The dataset is downloadable from here https://doi.org/10.23729/7b22c271-5e25-40fe-aa6a-f5d0330b1872.

Install dependencies

pip install -r requirements.txt

Preprocess the dataset

The dataset must be preprocessed with preprocess_tfrecords.py script before training the model. Change /path/to in the filepaths to the actual location where you want to store the files.

python preprocess_tfrecords.py --dataset_stats /path/to/stats.json --output_dir /path/to/processed/train/ input_file.tfrecord.GZIP

You might want to use GNU parallel or similar to parallellize the processing e.g.

# assuming that train_files.txt includes filepaths to all files you want to preprocess
parallel --halt now,fail=1 -j 45 -I{} python preprocess_tfrecords.py --dataset_stats /path/to/stats.json --output_dir /path/to/processed/train/ {} :::: train_files.txt

Do same for the validation data.

# assuming that val_files.txt includes filepaths to all files you want to preprocess
parallel --halt now,fail=1 -j 45 -I{} python preprocess_tfrecords.py --dataset_stats /path/to/stats.json --output_dir /path/to/processed/val/ {} :::: val_files.txt

Train the neural network

python main.py \
        --train_data "/path/to/processed/train/records/*.tfrecord.GZIP" \
        --val_data "/path/to/processed/val/records/*.tfrecord.GZIP" \
        --no_checkpoints \
        --epochs 50

The training script is logging the training metrics to ./logs directory. You can see the logs with tensorboard.

tensorboard --logdir logs/ --samples_per_plugin "images=50"

Epoch loss plot with the default parameters (blue = train, red = validation):

Computing the classifier accuracies

Use the trained model checkpoint to add predictions to the dataset

python generate_change_dataset_with_predictions.py \
        --model_checkpoint "logs/CHECK_THE_CORRECT_DIR/checkpoints/final" \
        --output_dir "/path/to/simulated-change-with-prediction/" \
        "/path/to/val/records/*.tfrecord.GZIP"

Compute the results

For threshold classifier:

python threshold_classifier.py  --save_filename minus2_5dB-shift.png --simulated_change_shift -2.5 "/path/to/simulated-change-with-prediction/*.tfrecord.GZIP"

And for SVC classifier:

python svm_classifier.py "/path/to/simulated-change-with-prediction/*.tfrecord.GZIP"

Simulated change

This repository includes the generate_change_dataset.py script that was used to generate the simulated change dataset for the experiments. However, the script is too tightly coupled with the database for it to be executable anywhere. The script requires access to the PostgreSQL database that stores the SAR image rasters, and the database is too large to be shareable. However, you can request the dataset that was used to run the experiments. The dataset includes the simulated changes for the validation samples.

Ablation Study

An ablation study was conducted for the different features to see what features are important for the model. The experiment can be repeated using the --ablation_study argument for the main.py script (e.g. python main.py --ablation_study snow_depth trains the neural network without the snow depth feature). The resulting validation loss plot from the different runs is interesting, but did not fit to the paper, therefore it is presented here. The loss clearly shows that the 'Orbit Direction' feature is the most important one, and without it the validation loss is clearly higher than when the other features are removed. 'Temperature', 'Precipitation', and 'Incidence Angle' are also important features. Removing the 'Snow Depth' and 'Satellite Id' features did not have much effect to the resulting validation loss as removing the other features.

Citation

@ARTICLE{10286479,
  author={Alatalo, Janne and Sipola, Tuomo and Rantonen, Mika},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Improved Difference Images for Change Detection Classifiers in SAR Imagery Using Deep Learning}, 
  year={2023},
  volume={61},
  number={},
  pages={1-14},
  doi={10.1109/TGRS.2023.3324994}}

Acknowledgements

This research was conducted in Jamk University of Applied Sciences and was funded in part by the Regional Council of Central Finland/Council of Tampere Region and European Regional Development Fund as part of the project Data for Utilisation -- Leveraging digitalisation through modern artificial intelligence solutions and cybersecurity (grant number A76982), in part by the REACT-EU Instrument as part of the European Union’s response to the COVID-19 pandemic as part of the project coADDVA - ADDing VAlue by Computing in Manufacturing (grant number A77973), and in part by the European Union and the Regional Council of Central Finland with the Just Transition Fund as part of the project Finnish Future Farm (grant number J10075).