Unet-diff deforestation model inference

Build image

For building image locally you need to place model into
models/unet-diff.pth

And also download S2A_OPER_GIP_TILPAR_MPC__20151209T095117_V20150622T000000_21000101T000000_B00.kml(link) file and place it into data/landcovers/S2A_OPER_GIP_TILPAR_MPC__20151209T095117_V20150622T000000_21000101T000000_B00.kml

then run git submodule update --init --recursive

and finally

docker build -t quantumobile/deforestation .

Docker run command

docker run \
    --runtime nvidia \
    -e "AOI=POLYGON ((-85.299088 40.339368, -85.332047 40.241477, -85.134979 40.229427, -85.157639 40.34146, -85.299088 40.339368))" \
    -e "START_DATE=2020-05-01" \
    -e "END_DATE=2020-06-30" \
    -e "SENTINEL2_GOOGLE_API_KEY=/input/sentinel2_google_api_key.json" \
    -e "SENTINEL2_CACHE=/input/SENTINEL2_CACHE" \
    -e "OUTPUT_FOLDER=/output" \
    -v path_to_SENTINEL2_CACHE:/input/SENTINEL2_CACHE \
    -v path_to_sentinel2_google_api_key.json:/input/sentinel2_google_api_key.json \
    -v path_to_/output:/output \
    quantumobile/deforestation

Push to repository

docker push quantumobile/deforestation

Pull image from repository

docker pull quantumobile/deforestation

Forest Change Detection

Description

This is a source code repository for DEEP LEARNING FOR REGULAR CHANGE DETECTION IN UKRAINIAN FOREST ECOSYSTEM WITH SENTINEL-2 (Kostiantyn Isaienkov, Mykhailo Yushchuk, Vladyslav Khramtsov, Oleg Seliverstov), 2020.

Repository structure info

baseline - scripts for deforestation masks predictions with baseline models
time-dependent - scripts for forest change segmentation with time-dependent models, including Siamese, UNet-LSTM, UNet-diff, UNet3D models

Setup

All dependencies are given in requirements.txt Main setup configuration:

python 3.6
pytorch==1.4.0
torchvision==0.5.0
catalyst==19.05
segmentation-models-pytorch==0.1.0

Tested with Ubuntu + Nvidia GTX1080ti with Cuda==10.1 and Nvidia GTX1060 with Cuda==10.2. CPU mode also should work, but not tested.

Dataset

You can download our datasets directly from Google drive for the baseline and time-dependent models. The image tiles from Sentinel-2, which were used for our research, are listed in tiles folder.

The data include *.geojson polygons:

baseline: 2318 polygons, 36UYA and 36UXA, 2016-2019 years;
time-dependent: 36UYA (two sets of separated annotations, 278 and 123 polygons -- for spring and summer seasons respectively, 2019 year) and 36UXA (1404 polygons, 2017-2018 years). The files contain the following columns: tileID (ID of a tile, which was annotated), img_date (the date, at which the tile was observed), and geometry (polygons of deforestation regions).

Training

Reproduce results

To reproduce the results, presented in our paper, run the pipeline (download data, prepare images, train the models), as described in README files inbaseline and time-dependent folders.

Training with new data

To train the models with the new data, you have to create train/valid/test (*.csv) files with specified location of images and masks, and make a minor changes in Dataset classes (for more information about location of these classes, see README files in baseline and time-dependent folders).

Citation

If you use our code and/or dataset for your research, please cite our paper:

@ARTICLE{Isaienkov2021,
  author={K. {Isaienkov} and M. {Yushchuk} and V. {Khramtsov} and O. {Seliverstov}},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  title={Deep Learning for Regular Change Detection in Ukrainian Forest Ecosystem With Sentinel-2},
  year={2021},
  volume={14},
  number={},
  pages={364-376},
  doi={10.1109/JSTARS.2020.3034186}
}

QuantuMobileSoftware/forest_change_detection