Blog – Geospatial entities as a key for unlocking value adding data fusion models of the next generation

Domen Mongus, EUROGI Executive Committee member 

13/03/2018

While we are all becoming aware of recent rapid growth of data sources and streams, the hype of Big Data is close to an end. In other words, we are slowly moving away from hopes and dreams as we are entering into a phase of stable technology development, with already emerging established solutions. Nevertheless, by focusing on high-performance cloud computing infrastructures, it seems like we have primarily addressed the volume and velocity aspects of Big Data in the past, while dealing with data variety remains a fairly intact challenge. Within a veritable jungle of available data sources, types, and data structures, this is somehow expected as variety is arguably the most difficult aspect to addressed. Still, although being somewhat critical of progress towards data driven decision making, new buzzwords related to data fusion, high-level feature extraction, and feature learning are already looming on the horizon, particularly within the domain of spatiotemporal analytics.
From this perspective, geospatial technologies offer several key advantages over related technologies, as for example the wide acceptance of standardised data exchange and data manipulation APIs simplifies integration of system to the level that is unimaginable in most of the other ICT fields. Geographic information systems are, consequently, now merging with various applications outside of land management and other traditional GIS domains, as spatial concepts are being brought into in the gaming industry, entertainment and augmented reality, autonomous driving, and position awareness of smart devices in general. But perhaps the most interesting and up to now significantly underutilized aspect concerns data in the emerging domain area of data fusion. Geospatial entities, may it be points, lines or polygons, allow for linking of heterogeneous data sources and streams at a semantic level. Data structuring concepts with an underlying spatial basis may indeed become critical in new architectures of artificial intelligence, as everything happens at a certain place in a certain time. While different sensory systems may be used to measure the events, a standardised notation of location and time provides a way to interlink the measurements in the form of spatially embedded graphs (i.e. nodes attributed with geolocation and liked by weighted edges that allow for modelling their relation in the form of activation functions). Building machine learning algorithms then becomes relatively easy as data representation with a spatial basis naturally resembles a neural network itself, while interlinked sensors may be of physical or of social sensing types.
The potentials of such architectures with a spatial aspect, therefore, do not only concern capturing and/or modelling earth-related phenomena through IoT systems for spatial awareness, but also allow for integrating social aspects through the on-the-fly fusion of open-data sources and ever more active social sensors. While machine learning applied on top of that should result in accurate prediction models, these models may indeed contain structured representation of knowledge that has long remained hidden due to the complexity of the underlying spatiotemporal relations. We may, therefore, speculate that it will soon become more effective to analyse prediction models themselves rather than the data and, thus, triggering the ages of true machine-based knowledge discovery.