The Framework
How is AI supporting mapping of natural capital?
Accurate and reliable maps of natural capital are vital for underpinning investment in, and management of, nature restoration projects. Maps of natural capital are needed at a scale that is prohibitive to collect through fieldwork alone. AI is increasingly being used to scale up traditional assessments of natural capital by leveraging AI models to interpret indirect measurements, such as satellite imagery or lidar. Examples of AI derived maps include soil carbon stocks, tree canopy height, and national or global landcover maps.
How are these maps created?
There is a myriad of AI algorithms, datasets and methodologies that can be used to map natural capital, so it is difficult to generalise. However, these maps will usually involve:
- 1.Mapping data that is available at scale, such as satellite imagery, aerial lidar or thematic mapping. Increasingly, data from multiple sources and dates are combined to characterise the Earth more completely.
- 2.An AI model that interprets the mapping data and creates the output map. This can range from simple machine learning models, such as random forests or KNN classifiers, to deep learning models such as YOLO or DINOv2.
- 3.Data to train the model and validate its predictions. These could be field observations, such as tree height measurements, or could be generated by human interpretation of imagery e.g. landcover labels.
What are the challenges for trusting AI derived maps?
Conversations with map users conducted during the TaiM project revealed several challenges for trusting AI derived maps that were leading to reduced uptake of these products. For example:
- •Characteristics vary greatly product to product, meaning a lot of effort is required to evaluate fitness for purpose.
- •Full methodologies are rarely shared due to IP considerations, especially for products created by commercial organisations.
- •Accuracy is usually assessed against a subset of training data, meaning errors in the training data will pass through to the accuracy analysis.
- •Accuracy analysis usually only considers the final map outputs; sources of error, such as those shown in the diagram below, are often not considered.
How does TAiM help?
We believe that the best approach to validation is to compare maps outputs against independent assessments of ground conditions. However, for this comparison to be fair it must be done on the terms of the map, i.e. at the same spatial resolution and using the same measurement / classification schema. TaiM offers:
- 1.High-accuracy field data for a range of natural assets collected across multiple Scottish field sites.
- 2.A map accuracy framework that supports fair comparison between field data and map outputs.
- 3.A portal for map validation that requires a map output that covers one of our field sites, and the ability to answer some questions about the map creation methodology.
- 4.Standardised reports and maps that show where, and why, errors have occurred.