Deep learning algorithms appear as a major breakthrough in GIS scope: neural networks are able to do semantic segmentation on aerial images, so as to identify building footprints, roads, and so on.

Oslandia is an opensource company studying and exploiting geospatial data, with an extensive R&D activity about geospatial data science. This presentation will detail some of our Python routines in terms of geospatial data handling.

We will describe our processes from raw data to prediction results. As the main step of the pipeline, machine learning techniques (e.g. convolutional neural networks for image semantic segmentation with Keras) produce valuable predictions. In the case of geospatial data, a postprocessing step is often necessary for displaying the results in web applications and GIS tools.

A concrete illustration of our results will be provided through a light Flask application designed for demonstration purpose.

Taking advantage of the emergence of Open Data platforms dedicated to urban services, one can try to understand the functioning of cities. The biggest challenge is no more to get the data; but to structure it, analyze it, extract new information from it, and design clever representations in order to visualize it.

This presentation will focus on a recent open source study made by Oslandia about bike-sharing systems in France. Our dev stack is largely based on Python tools, from back-end (a data pipeline designed with Luigi) to front-end (with Flask API and web application).

The most important data processing steps will be detailed, and a particular attention will be paid to inherent machine learning problems, like bike stand classification, or bike availability prediction. To that matter, we target a better comprehension of urban areas, and value creation for bike-sharing system users.

A live demo of the web application will end the presentation.