Large Scale High-Resolution Land Cover Mapping with Multi-Resolution Data

Repo accompanying the CVPR 2019 paper "Large Scale High-Resolution Land Cover Mapping with Multi-Resolution Data".

Data

The dataset and its documentation can be found at LILA.science. The download_all.sh script will download and unzip a copy of this dataset to the ./chesapeake_data/ directory. Note: this script assumes that the azcopy program is on your PATH; azcopy can be downloaded here.

Requirements

Pre-trained keras models were generated with Python 3.6. See requirements.txt for the library versions that we used when developing. Generally, these libraries are needed:

• numpy
• pandas
• tensorflow-gpu
• Keras
• segmentation-models
• shapely
• rasterio

References

Please cite the following papers if you use this work:

@inproceedings{robinson2019large,
  title={Large Scale High-Resolution Land Cover Mapping With Multi-Resolution Data},
  author={Robinson, Caleb and Hou, Le and Malkin, Kolya and Soobitsky, Rachel and Czawlytko, Jacob and Dilkina, Bistra and Jojic, Nebojsa},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2019},
  url={http://openaccess.thecvf.com/content_CVPR_2019/html/Robinson_Large_Scale_High-Resolution_Land_Cover_Mapping_With_Multi-Resolution_Data_CVPR_2019_paper.html}
}

@inproceedings{malkin2018label,
  title={Label super-resolution networks},
  author={Malkin, Kolya and Robinson, Caleb and Hou, Le and Soobitsky, Rachel and Czawlytko, Jacob and Samaras, Dimitris and Saltz, Joel and Joppa, Lucas and Jojic, Nebojsa},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2019},
  url={https://openreview.net/forum?id=rkxwShA9Ym},
}

Todo

• Change train_model_landcover.py to save models without the superres loss, jaccard loss, or Lambda layers as these can cause problems with saving/loading in different versions.
• Re-write eval_landcover_results.sh in Python.
• Create a test script that computes accuracy on the fly without saving model results.
• Make it easy to target other label sets (e.g instead of 4-class classification).