phil1108 / gsis

German Satellite Image Segmentation

GSIS

German Satellite Image Segmentation

German Satellite Image Segmentation - The smart way to get Training Data for aerial Image Classification


View More Examples

Table of Contents

• About the Project
• Getting Started
  • Installation
  • Download
• Usage as DNN Input
• Acknowledgments

About The Project

There are many datasets available for segmentation of aerial images. However most of them lack of scale, detail or are under commercial license. This projects aims to:

• Provide a large amount of training data (35.000 km²) fully annotated with segments
• Consists of a vast amount of different classes (Streets, Vegetation, Homes, Garages, Paths ...)
• Covers rural and urban areas
• Helps you create your individual datasets with adoptable parameters and extent
• Allows you to train an CNN to detect estate boundaries

If you wanna try out an example just download one preformed database (size of around !!! GB) at kaggle.

The code itself is released under the MIT License however the data is provided by the federal state of North Rhine-Westphalia under the Data licence Germany – attribution – version 2.0.

Getting Started

To create or adpot the databases you need one of the following system setups:

1. Docker where you can just run the prebuild docker files (the easiest solution)

2. OSGeo4W (for Windows) or OSGeolive where nearly all Installation requirements are already build in

3. Python3 (Tested with Ubuntu 18.04 but Windows probably may also work) where you have to install all packages from Installation

Installation

Download

Usage as DNN Input

For each .jp2 Aerial Image the according Image segmentation is computed and saved as .png file. In Standard settings these are binary files with 0 for no Object and 255 representing an Object (default any type of house) Please note that the .jp2 have an additional Infrared Band.

Acknowledgments

xmi2db library is used in this project