Matt Lythe, Managing Director
Some might say AI or more specifically machine learning is nothing new to Geographic Information Systems (GIS). They are partly correct as algorithms like cluster analysis and regression have been used for decades by GIS professionals. However, machine learning can also take geospatial analysis to new heights.
In this post I will explore the potential for GIS data to train and build better machine learning models. I will look at some practical examples of this and discuss what lies ahead in the field of artificial intelligence.
A GIS contains a wealth of information classified by location. These geographical “systems of record” make for high quality training inputs for machine learning models. In particular - structured geospatial data such as land parcels, roads, land use and asset locations are excellent resources to guide and inform machine learning models.
Machine learning, at its core, is composed of approaches that learn from data rather than being explicitly programmed. Deep learning—so called because it uses “deep” artificial neural networks that are composed of hundreds of connected layers of algebra has enabled something of a revolution in the processing of structured and unstructured data of all types – including geospatial. For those working in GIS, this means new solutions are now available to address problems previously considered too resource intensive or costly.
Deep learning has created the ability for example to accurately detect objects and categorise pixels from imagery at scale. Neural networks can be trained to detect and document geometries of buildings, roads, crops, excavated land… or to detect and classify assets and their condition.
Going one step further, with targeted training, deep learning can also detect and generate new attributes against these geometries for example condition of buildings or surface material.
The use of artificial intelligence in conjunction with GIS has already had a big impact on location centric applications such as traffic monitoring and management, micro mobility services and ride sharing.
More recently, some of the less structured data captured by organisations such as raster data, video, voice and text are now being used by machine learning models to generate insights and predictions. At Lynker Analytics we have been applying and refining these techniques in the detection and classification of a wide range of target classes across multiple industries including marine, conservation, agriculture and transportation. Check out some of our recent blog posts to earn more about this work.
So - in the sub-field of computer vision it is true to say that AI systems are benefiting greatly from GIS data streams while in turn geographic information systems themselves are the direct beneficiaries of machine learning by way of generating data with higher fidelity, frequency and at a lower cost.
What about in the field of predictive analytics – where there is in all probability more business and operational value for organisations?
In this domain GIS users have been using AI for some time. Regression analysis, kriging, hot spot analysis and clustering have been in use for decades. But, we’re still in the early days of applying more complex machine learning models such as Deep Neural Networks and Gradient Boosted Machines within a fully geospatial context.
Machine learning researchers and data scientists haven’t necessarily understood the inference already inherent within a GIS that could lead to powerful new inputs and feature engineering for machine learning models. The geographic relationships between objects have often been ignored or handled poorly e.g. simply tagging with a latitude, longitude. Now we can turbo boost our machine learning models with geospatial analytics.
Rather than acting alone, an AI system that uses the inter-connected feature relationships held within a GIS will potentially perform better. An AI when applied within a spatial context may help us understand why assets or systems that work in one place fail in another. It will work well when processing large volumes of structured data such as observations from instruments in the field, or attributes from a feature layer and understanding their influence on a situation.
Applications of this joined up approach include predicting the probability of congestion or accidents in transport networks, water quality or vegetation condition across differing land use classes through to forecasting sales in a retail setting. It can also help us understand the locational aspects that influence more difficult to identify associations that may have future consequences. Programmes like Microsoft’s AI for Earth is a good example of where these intersecting technologies are now delivering new insights and value
The application of AI within a spatial context is just getting started. It is not about producing human level intelligence – that is some way off - but it is fast becoming apparent that these connected tools when used in unison can create better decision systems.