Skip to content

Latest commit

 

History

History
115 lines (77 loc) · 8.8 KB

README.md

File metadata and controls

115 lines (77 loc) · 8.8 KB
Raw

Deep learning for Earth Observation

http://www.onera.fr/en/dtim https://www-obelix.irisa.fr/

This repository contains code, network definitions and pre-trained models for working on remote sensing images using deep learning.

We build on the SegNet architecture (Badrinarayanan et al., 2015) to provide a semantic labeling network able to perform dense prediction on remote sensing data.

Motivation

Earth Observation consists in visualizing and understanding our planet thanks to airborne and satellite data. Thanks to the release of large amounts of both satellite (e.g. Sentinel and Landsat) and airborne images, Earth Observation entered into the Big Data era. Many applications could benefit from automatic analysis of those datasets : cartography, urban planning, traffic analysis, biomass estimation and so on. Therefore, lots of progresses have been made to use machine learning to help us have a better understanding of our Earth Observation data.

In this work, we show that deep learning allows a computer to parse and classify objects in an image and can be used for automatical cartography from remote sensing data. Especially, we provide examples of deep fully convolutional networks that can be trained for semantic labeling for airborne pictures of urban areas.

Content

Deep networks

We provide a deep neural network based on the SegNet architecture for semantic labeling of Earth Observation images. The network is separated in two files :

  • The network definition (the architecture) in the prototxt format (for the Caffe framework),
  • Pre-trained weights on the ISPRS Vaihingen dataset and ISPRS Potsdam datasets.

All the pre-trained weights can be found on the OBELIX team website (backup link).

Examples :

Data

Our example models are trained on the ISPRS Vaihingen dataset and ISPRS Potsdam dataset. We use the IRRG tiles (8bit format) and we build 8bit composite images using the DSM, NDSM and NDVI. The ground truth files are color-encoded and should be converted to the numerical labels, e.g. {0,1,2,3,4,5} instead of {[255,255,255],[0,0,255],[0,255,255],[0,255,0],[255,0,255],[255,0,0]} using the convert_gt.py script.

You can either use our script from the OSM folder (based on the Maperitive software) to generate OpenStreetMap rasters from the images, or download the OSM tiles from Potsdam here.

Jupyter notebooks

In addition to our models, we provide several Jupyter notebooks to :

  • pre-process remote sensing images and build a dataset (with the example of the ISPRS Vaihingen dataset)
  • train a deep network for semantic labeling of remote sensing images
  • apply a deep network to label automatically remote sensing data

Configuration variables are grouped in the config.py file. This the one you have to edit to suit your needs. The scripts that can be used are :

  • Extract images : use a sliding window to divide high resolution images in smaller patches
  • Create LMDB : convert the small patches into LMDB for faster Caffe processing
  • Training : train a model on the LMDB

Those scripts are available either using Python or Jupyter. Two inference scripts are available but only for Python (though they can easily be imported in Jupyter).

How to start

This an example of how to start using the ISPRS Vaihingen dataset. This dataset contains orthoimages (top/ folder) and ground truthes (gts_for_participants/) folder.

  1. First, we need to edit the config.py file. BASE_DIR, DATASET and DATASET_DIR are used to point to the folder where the dataset is stored and to specify a unique name for the dataset, e.g. "Vaihingen". label_values and palette define the classes that will be used and the associated colors in RGB format. folders, train_ids and test_ids define the folder arrangement of the dataset and the train/test split using unique numerical ids associated to the tiles.
  2. We need to transform the ground truth RGB-encoded images to 2D matrices. We can use the convert_gt.py script to do so, e.g. : python convert_gt.py gts_for_participants/*.tif --from-color --out gts_numpy/. This will populate a new gts_numpy/ folder containing the matrices. Please note that the folders value for the labels should point to this folder (gts_numpy/).
  3. Extract small patches from the tiles to create the train and test sets : python extract_images.py
  4. Populate LMDB using the extracted images : python create_lmdb.py
  5. Train the network for 40 000 iterations, starting with VGG-16 weights and save the weights into the trained_network_weights folder : python training.py --niter 40000 --update 1000 --init vgg16weights.caffemodel --snapshot trained_network_weights/
  6. Test the trained network on some tiles : python inference.py 16 32 --weights trained_network_weights/net_iter_40000.caffemodel

Requirements

You need to compile and use our Caffe fork (including Alex Kendall's Unpooling Layer) to use the provided models. Training on GPU is recommended but not mandatory. You can download the fork by cloning this repository and executing :

# Clone this repository
git clone https://github.com/nshaud/DeepNetsForEO.git

cd DeepsNetsForEO/

# Initialize and checkout our custom Caffe fork (upsample branch !)
git submodule init

git submodule update

Our Caffe version will be available in caffe folder. You can then follow with the usual compilation instructions.

References

If you use this work for your projects, please take the time to cite our ACCV'16 paper :

https://arxiv.org/abs/1609.06846 Nicolas Audebert, Bertrand Le Saux and Sébastien Lefèvre, Semantic Segmentation of Earth Observation Data Using Multimodal and Multi-scale Deep Networks, Asian Conference on Computer Vision, 2016.

@inproceedings{audebert_semantic_2016,
	title = {Semantic {Segmentation} of {Earth} {Observation} {Data} {Using} {Multimodal} and {Multi}-scale {Deep} {Networks}},
	url = {https://link.springer.com/chapter/10.1007/978-3-319-54181-5_12},
	doi = {10.1007/978-3-319-54181-5_12},
	language = {en},
	urldate = {2017-03-31},
	booktitle = {Computer {Vision} – {ACCV} 2016},
	publisher = {Springer, Cham},
	author = {Audebert, Nicolas and Le Saux, Bertrand and Lefèvre, Sébastien},
	month = nov,
	year = {2016},
	pages = {180--196}
}

License

Code (scripts and Jupyter notebooks) are released under the GPLv3 license for non-commercial and research purposes only. For commercial purposes, please contact the authors.

https://creativecommons.org/licenses/by-nc-sa/3.0/ The network weights are released under Creative-Commons BY-NC-SA. For commercial purposes, please contact the authors.

See LICENSE.md for more details.

Acknowledgements

This work has been conducted at ONERA (DTIM) and IRISA (OBELIX team), with the support of the joint Total-ONERA research project NAOMI.

The Vaihingen data set was provided by the German Society for Photogrammetry, Remote Sensing and Geoinformation (DGPF).