AI & Machine Learning for monitoring Nature Based Solutions is one of the core skills we have developed at CSX says Mark Hopper, CSX’s technical team lead.
At CSX we have developed and are now advancing Machine Learning algorithms to identify peatland erosion features and individual trees using RGB imagery collected using drones. The importance of using drones for the data gathering is due to the high resolution imagery that can be collected with drones.
Drones vs Satellite Data
The open source satellite data we can all access, such as Sentinel 2 from ESA, has a resolution of 10m per pixel. The paid for satellite data is generally to a resolution of around 2m, whilst the 50cm resolution satellite data is cost prohibitive to currently be commercially viable to monitor nature based solutions.
Using drones we are able to gather imagery data at the resolutions we have found necessary for accurate interpretation through the application of AI. For trees this is 2.5cm GSD per pixel, peatlands at 1.5cm and Biodiversity Net Gain (BNG) at 0.75cm GSD. We have even now got to the point of 0.03cm GSD to test how we can improve our quality assurance processes.
How we make ML models for AI
The starting point for development and training of our Artificial Intelligence is through image annotation. This Ground Truth data being the annotated imagery from which the ML algorithm learns the features we want to identify. Once the ML system has been trained we can move on to accelerate our creation of Ground Truth data and train more precise models for identifying trees and their characteristics. One such method to do this we are exploring is utilising the Canopy Height Model (CHM) created from the imagery.
These characteristics are used to assess the volume of Carbon stored within a woodland. This calculation being ground-truthed using the outputs from the Quantitative Structure Model (QSM) we run on the processed point cloud collected using our Terrestrial Laser Scanner (TLS). For more on this see our blog post “Surveying landscapes with LiDAR sensors and Terrestrial Laser Scanners”.
The peatland Machine Learning uses a similar methodology to the woodland Machine Learning. It is trained using Ground Truth data, where we have annotated common erosion features found on peatland, including grips and gullies. This has then enabled CSC to produce a scope of work for the restoration of peat bogs using only a drone flight.
To increase my knowledge and skill base with regards to AI & Machine Learning for monitoring Nature CSX have enrolled me onto the Teesside University Skills Bootcamp. Further to this initial involvement I’m delighted that we’ve recently won a Knowledge Transfer Partnership bid with Teesside University to further expand our AI development resources and knowledge.
We are excited about the opportunities for expanding our AI capabilities, both within deep learning and identifying different Machine Learning algorithms and when these are best applied. These algorithms broadly differ by the level of supervision required. Some are “Supervised” algorithms which require ground truth data to know what features/patterns they are looking for. Others are “Unsupervised” algorithms which look for any patterns and features within a dataset. Unsupervised algorithms can be advantageous as it doesn’t require you to know about the pattern beforehand and they can find relationships you may not expect to find!
CSX’s machine learning running a tree count and measurement on a section of conifer woodland:
