Here you will find links to Odyssey Geospatial's blog posts, open source software, and R&D projects.
Positional encoding is a key technology enabling Geospatial AI and ML. The idea is that one can create vector encodings for shapes -- Points, LineStrings, and Polygons -- in a way that captures their key geometric properties. Those vectors then serve as input to ML and AI models, making the whole range of modern data processing technology applicable for vector-format geospatial data. This is a series of blog posts focusing on a method called Multi-Point Proximity encoding.
This post introduces the idea of Multi-Point Proximity encodings, and shows that they capture key geometric properties of the objects that they encode. Link
This post goes into more detail on the properties of MPP encodings: they are continuous, they exhibit desirable types of invariance, and they capture information that can be used for decoding. Link
This post shows that one can estimate pairwise spatial relationsbhips from MPP encodings: containment, overlap, and so on. Link
An example of applying geometric encodings in a realistic example use case: estimation of earthwquake rates based on the local distribution of geological features. This builds a neural nmetwork whose inputs are entirely derived from vector-mode geiospatial data. Link
We are actively working on new research, and will post here when ready.
These are installable python packages useful for various Geospatial / AI / ML tasks
Research notebooks used in R&D on geospatial encodings
This a github repo with a series of notebooks illustrating how geospatial encodings can capture pairwise spatial relationships.
We use cookies to analyze website traffic and optimize your website experience. By accepting our use of cookies, your data will be aggregated with all other user data.