Skip to content

Latest commit

 

History

History
86 lines (61 loc) · 4.01 KB

README.md

File metadata and controls

86 lines (61 loc) · 4.01 KB

Land cover translation with an assymetrical-Unet

This repo is the official implementation of the paper "Toward a yearly country-scale CORINE Land-Cover map without using images: a map translation approach".

Table of Contents:

Reference:

Install

For conda users :

conda env create -f environment.yml

For pip users : open with a text editor the environment.yml and install manually the packages.

Dataset

Datasets can be downloaded from the following zenodo archive.

The folder holds two dataset :

  • the first one is the dataset presented in the paper (oso-to-clc)
  • the second one contains partially self-covering-patches (oso-to-clc-self-covering) in order to make possible to create a France wide map without patches-edge effects. Thus it should not be used in a training procedure. Example of result can be observe on this website.

Folder structure is presented below :

├── oso_to_clc
|  └── train
|  |  |  └── oso_2018
|  |  |  └── oso_2016
|  |  |  └── clc _2018
|  |  |  └── clc_2012
|  └── test
|  └── val
|  └── metadata.json
|
├── oso_to_clc_self_covering
|  └── full
|  |  |  └── oso_2018
|  |  |  └── oso_2016
|  |  |  └── clc _2018
|  |  |  └── clc_2012
|  └── metadata.json
|
├── ground_truth.gpkg

Useful consideration :

  • The manually build ground_truth used in this study is made available in a geopackage fromat.
  • The metadata.json holds the localisation of each patch.
  • Each patch is encoded in a numpy binary format (.npy)

Tutorial:

This github repository follow the structure proposed Hager Rady and Mo'men AbdelRazek. For more information on the repository structure, and how to handle launch of the code refers to this github repo.

Once the python environnement is correctly set. Download the dataset and place it in the data folder. Then launch the run.sh script for a simple default train with default features in scenario 1 setup. Change the config file path in the run.sh file for scenario 2 or scenario 3.

Algorithm parameters are stored in the configs/ folder. By default parameters are set to run on computer with low ram and GPU capacity. On a powerful machine we recommend to adapt the following parameters :

  • full test : False -> True . By default statistic in test phase are only computed on 6022 ground truth. If set to true, comparison between prediction and CLC will also be computed (agreement between prediction and CLC 2018, EPI). Note that this is option should only be set to true with at least 40 Go RAM.
  • train_batch_size,test_batch_size,valid_batch_size : 4 ->n. batch size can be adapted depending on the capacity of your computer. Note that results presented in the paper are computed with batch size = 24.

Once the network is trained, one just have to adapt the create_france_wide_map.sh with the desired model and scenario to produce a France wide map (the script uses the oso_to_clc_self_covering dataset). Be aware that in this setup the result are displayed both on seen and unseen patches during training (all France) thus the result is unsuitable for quantitative quality assessment.

Acknowledgement

License:

This project is licensed under MIT License - see the LICENSE file for details