For a few years, more and more open geospatial datasets have been released regarding aerial and satellite imagery. In parallel, a wide range of geospatial tools and softwares emerged in order to exploit this amount of data, in particular through image semantic segmentation. This kind of tools are based on Artificial Intelligence techniques, especially convolutional neural networks.

At Oslandia, we proposed Deeposlandia [1] to address such a point. By considering a set of (high-resolution) images, one may easily know the composition of the image at the pixel level. This opens the door for use cases like building footprint recognition, as an example.

This presentation will depict how to go further and bridge the gap between AI-related softwares and QGIS. Our main focus will be to introduce QDeepLandia [2], a QGIS plugin that aims at providing basic semantic segmentation features for GIS users. AI techniques are composed of two different steps: a long resource-intensive training step and a quicker inference steps. While the former does not aim at figuring into a desktop application like QGIS (a huge amount of involved images, specific GPU resources), the latter suits the needs of users who want to analyze rasters and produce vectorized segmentation results on-the-fly.

Hence after presenting some considerations about state-of-the-art regarding semantic segmentation facilities in QGIS, the presentation will focus on the technical locks which must be overcome in terms of dependency management and packaging. Finally a presentation of Deeposlandia and QDeepLandia will be done.

[1] https://gitlab.com/Oslandia/deeposlandia
[2] https://gitlab.com/Oslandia/qgis/QDeeplandia