Chihyulai/robosat.pink

Semantic Segmentation ecosystem for GeoSpatial Imagery

RoboSat.pink

Semantic Segmentation ecosystem for GeoSpatial imagery

Purposes:

• DataSet Quality Analysis
• Change Detection highlighter
• Features extraction and completion

Main Features:

• Provides several command line tools, you can combine together to build your own workflow
• Follows geospatial standards to ease interoperability and data preparation
• Build-in cutting edge Computer Vision model and loss implementations (and allows to replace by your owns)
• Support either RGB or multibands imagery (as multispectral or hyperspectral)
• Allows Data Fusion (from imagery or rasterized vectors)
• Web-UI tools to easily display, hilight or select results
• High performances

Documentation:

Tutorials:

• How to use plain OpenData to create a decent training OpenDataSet

Config file:

• RoboSat.pink configuration file

Tools:

• rsp cover Generate a tiles covering, in csv format: X,Y,Z
• rsp download Downloads tiles from a remote server (XYZ, WMS, or TMS)
• rsp extract Extracts GeoJSON features from OpenStreetMap .pbf
• rsp rasterize Rasterize vector features (GeoJSON or PostGIS), to raster tiles
• rsp subset Filter images in a slippy map dir using a csv tiles cover
• rsp tile Tile raster coverage
• rsp train Trains a model on a dataset
• rsp export Export a model to ONNX or Torch JIT
• rsp predict Predict masks, from given inputs and an already trained model
• rsp compare Compute composite images and/or metrics to compare several XYZ dirs
• rsp vectorize Extract simplified GeoJSON features from segmentation masks

Presentations slides:

• @PyParis 2018
• @FOSDEM 2019

Installs:

With PIP:

pip3 install RoboSat.pink                                     # For latest stable version

or

pip3 install git+https://github.com/datapink/robosat.pink     # For current dev version

With Conda, using a virtual env:

conda create -n robosat_pink python=3.6 && conda activate robosat_pink
pip install robosat.pink                                      # For latest stable version        

With Ubuntu 18.04, from scratch:

sudo sh -c "apt update && apt install -y build-essential python3-pip"
pip3 install RoboSat.pink && export PATH=$PATH:~/.local/bin
wget http://us.download.nvidia.com/XFree86/Linux-x86_64/418.43/NVIDIA-Linux-x86_64-418.43.run 
sudo sh NVIDIA-Linux-x86_64-418.43.run -a -q --ui=none

With CentOS 7, from scratch:

sudo sh -c "yum -y update && yum install -y python36 wget && python3.6 -m ensurepip"
pip3 install --user RoboSat.pink
wget http://us.download.nvidia.com/XFree86/Linux-x86_64/418.43/NVIDIA-Linux-x86_64-418.43.run 
sudo sh NVIDIA-Linux-x86_64-418.43.run -a -q --ui=none

NOTAS:

• Requires: Python 3.6 or later
• GPU is not strictly mandatory, but rsp train would be -that- slower without.
• To test RoboSat.pink install, launch in a terminal: rsp -h
• Upon your pip PATH setting, you may have to update it: export PATH=$PATH:.local/bin
• PyTorch release published on PyPI is binded with CUDA 9. For CUDA 10, grab a wheel from PyTorch site.

DataSet:

• Training and validation datasets have to be tiled, using XYZ tiles format.
• A Dataset directory, so containing XYZ tiles, can be split as:

dataset
├── training
│   ├── images
│   └── labels
└── validation
    ├── images
    └── labels

• Tiles images formats could be any able to be read by GDAL.
• Tiles labels are expected to be PNG with single band.
• Tools producing XYZ tiles directory, generate also a web map client, for visual inspection.

Data Preparation:

Several ways to create your own training dataset, upon input data type:

NOTA: several inputs connected to a single arrow point means a logical OR (e.g. WMS or XYZ or TMS).

Architecture:

RoboSat.pink use cherry-picked Open Source libs among Deep Learning, Computer Vision and GIS stacks.

Related resources:

• Historical MapBox RoboSat github directory
• Christoph Rieke's Awesome Satellite Imagery Datasets
• Satellites in Global Development

Bibliography:

• U-Net: Convolutional Networks for Biomedical Image Segmentation
• Deep Residual Learning for Image Recognition
• Angiodysplasia Detection and Localization Using Deep Convolutional Neural Networks
• TernausNetV2: Fully Convolutional Network for Instance Segmentation
• The Lovász-Softmax loss: A tractable surrogate for the optimization of the IoU measure in neural networks
• Joint Learning from Earth Observation and OpenStreetMap Data to Get Faster Better Semantic Maps
• Albumentations: fast and flexible image augmentations

Contributions and Services:

• Pull Requests are welcome ! Feel free to send code... Don't hesitate either to initiate a prior discussion via gitter or ticket on any implementation question. And give also a look at Makefile rules.

• If you want to collaborate through code production and maintenance on a long term basis, please get in touch, co-edition with an ad hoc governance can be considered.

• If you want a new feature, but don't want to implement it, DataPink provide core-dev services.

• Expertise and training on RoboSat.pink are also provided by DataPink.

• And if you want to support the whole project, because it means for your own business, funding is also welcome.

Authors:

• Daniel J. Hofmann https://github.com/daniel-j-h
• Olivier Courtin https://github.com/ocourtin