Vicmap Machine Learning

Developing new and upgraded data products for Vicmap utilising machine learning with aerial imagery and LiDAR.   

The Challenge

The Victorian Department of Environment, Land, Water and Planning (DELWP) is the owner and custodian of Vicmap: the State’s authoritative foundational spatial data portfolio. The effort and resourcing to maintain Vicmap data is increasing at a rate of 5% to 10% per year. The maintenance is largely a manual process and to control costs, higher priority is given to high use products (e.g., property), meaning other products (e.g., vegetation) are not upgraded for long periods, resulting in out-of-date information being distributed to customers. The first of these projects aimed to prove the concept that Vicmap can be maintained and upgraded using intuitive data integration and machine learned feature extraction from DELWP’s extensive imagery and elevation data. The current project focuses on the Vicmap vegetation products, specifically individual tree locations for urban areas, along with a tree extent and density product for the whole state. The individual tree product will rely on an updated version of the first project’s prototype vegetation algorithm, while the tree extent and density products will rely on an alternative machine learning approach.

Partners

The Victorian Department of Environment, Land, Water and Planning (DELWP) was the primary partner in this project, and FrontierSI has been supported by Orbica.

The Solution

The scope of the first project was to prove the concept and performance of applying machine learning to Vicmap Imagery across urban, peri-urban, rural, and forested areas. The project was highly successful in delivering state-wide machine learning prototype algorithms that can identify and tackle the environmental differences in each landscape archetype.

The methodology adopted to develop and test the machine learning algorithms was:

  1. Human analysts create training sets over target areas.
  2. Develop prototype algorithm.
  3. Compare the output of the algorithm against the training sets.
  4. Compare with Vicmap specifications and requirements.

This first project developed three algorithms: one each for trees, buildings and swimming pools in sample areas. Through training and testing these algorithms, it was shown that the respective features can be automatically extracted with a higher degree of confidence than if defined by a human analyst. The tree algorithm was the most challenging and hence was heavily focussed on, using over 100,000 feature delineations to prove the potential of the technology for Vicmap. The detection rate for trees was better than 94% for the study area (however varied for different types and heights of vegetation); the swimming pool algorithm had a 99% detection rate; and although the building algorithm requires further development, it outperformed the current building footprints present in Vicmap Features of Interest and those available within GeoScape.

The success of the proof-of-concept project led to the current project, which aims to improve the Vicmap Vegetation products for the whole state. The prototype algorithm developed in the first project performed best at identifying individual trees in urban areas, which led to the decision to use it for an urban individual tree product, which will catalogue the location and heights of trees. This will allow for councils to monitor the growth of trees and manage these important assets. To better map the extent and density of tree coverage over the entire state, this project is developing a machine learning algorithm that detects tree cover on a ‘per pixel’ basis, resulting in a map of tree extent at 20cm resolution for the state, as well as a generalised density layer.

Impact

The success of the demonstrator project led to the current project and will ideally update the Vicmap Vegetation products once concluded. This will significantly enhance the detail, accuracy and compatibility of vegetation products to support analytics (i.e., the ability to spatially relate a holistic vegetation ledger, and map changes in vegetation over time).

More broadly, utilising aerial imagery and LiDAR to improve Vicmap will increase its value proposition for both customers and DELWP. Investment in machine learned feature extraction is expected to provide a significant economic benefit to the Victorian community. It will also allow DELWP to utilise the inherent value of its imagery and LiDAR archives to provide value to councils, and other customers of Vicmap, and theoretically unlock the untapped potential to create new data and analytics using Vicmap. Benefits of investing in machine learning include:

  • Timely and regular updates.
  • Improved operational efficiencies, productivity and sustainability.
  • Improved data provenance and quality.
  • Enable consumers to analyse changes over time effectively.
  • Map the ‘true’ positioning of features.
  • Developing 3D and 4D digital representation of Victoria.
  • Reduce costs.
  • Increase useability and reusability.
  • Build the value of the Vicmap brand.
  • Enable better evidence-based decision making.
Contact 

To learn more, contact FrontierSI at contact@frontiersi.com.au or Project Manager, Caitlin Adams, at cadams@frontiersi.com.au.