Mr Artiom Bondarenco1, Dr Andrew Higgins1, Dr Stephen McFallan1, Mr Adam McKeown1, Ms Caroline Bruce1, Mr Nick Smolanko1
1CSIRO, Brisbane, Australia
The Transport Network Strategic Investment Tool (TraNSIT) uses geospatial datasets from dozens of sources. Merging or conflating these datasets is critical for providing the range of variables required for the tool’s road and rail transport analysis. An accurate conflation of geospatial datasets is a classic GIS problem. Several approaches exist to perform vector to vector conflation, however they either do not work well with big data or require hard-coding of rules to determine the best fit between conflated features. This quickly can become extremely complex and time-consuming with big and diverse geospatial datasets, such as in the road network used for TraNSIT, containing over 3.2 million segments. We used a Deep Learning approach to conflate a vector dataset, containing classification data for heavy vehicle access, onto another vector dataset that is used for the modelling of vehicle movements in TraNSIT. A Deep Feed Forward Neural Network (DFFNN) was built using Keras package in Python. The DFFNN was trained on 80,360 manually conflated network segments and then prediction was made on a network dataset of 323,244 segments; both datasets comprised road related variables extracted from the HERE network. The resulting dataset represented truck classification with 94% accuracy. With further training and improvement of our DFFNN, it could serve as a powerful tool for a quick and accurate conflation of vector datasets.
Data analyst working in geospatial field.