It is unlikely that we will be able to directly collect all the data we require for future products; much of the most interesting data are about ‘invisible’ aspects of the real world such as socio-economic conditions or underlying environmental variables.
These must be inferred from the data that we can collect, such as imagery, Twitter feeds or customer activity. In essence we must reverse-engineer the real-world processes to predict variables that we are interested in from the collected data.
Real-world processes are very complex and it is rare that we can fully model them. Attempting to solve such ‘ill-posed’ problems with a manually-derived series of decisions and rules is inefficient and will probably result in a deficient solution. Machine learning offers a range of methods by which the relationship between the sensor data and the real-world variable of interest can be discovered automatically. The desired ‘reverse-engineering’ is achieved by optimising over a large dataset.
A research programme using machine learning is being established at Ordnance Survey. Initial work has investigated the potential to automatically populate products with quality measures and to derive new information from DSMs. Even more ambitious projects are being planned, including the discovery of regional character throughout GB and the development of a ‘representations’ database from which new products and enhancements can be rapidly derived.
For more information:
Contact: Anne Patrick, Research Project Coordinator