Collecting ground-truth data for Trustable AI in Mapping
Highlands Rewilding is excited be part of the consortium delivering the Trustable AI in Mapping (TAiM) project funded by Innovate UK, the UK’s innovation agency. The project has developed new standards to assess the reliability of Artificial Intelligence (AI) algorithms used in environmental mapping.
See our joint press release: From Black Box to Benchmark, Scottish Team Sets New Geospatial AI Standards for the Environmental Sector.
Our role in TAiM was to conduct detailed field surveys, providing robust data against which outputs from AI algorithms can be assessed, and to carry out research into the types and sources of ecological errors, asking, “Is ground-truth data ever really the truth?”.
Part of the project was a full National Vegetation Classification (NVC) survey of the Tayvallich Estate in mid-Argyll. This site includes a wide range of vegetation types and is a great example of a complex patchwork of habitats that pose a challenge for identification from satellite images.
TAiM enabled us to recruit three excellent Research Associates (Annabel Everard, Claudia Foster and Thorkil Osborne) to work alongside our Science team to develop our survey methodologies, carry out the surveys, and deliver the project. These detailed surveys covered four different habitat types: woodland, grassland, peatland and saltmarsh.
Woodland census survey
Left: Carrying out a woodland census survey
While habitat classification surveys provide a woodland type, for understanding biodiversity the species composition and structure, for example the variety of sizes of tree, is important. We carried out detailed woodland census surveys at Bunloit Estate (Inverness-shire), Beldorney Estate (Aberdeenshire) and Tayvallich Estate. Within the selected areas, we measured each tree, recording the species, height, diameter, whether it was multi-stemmed, and the presence of deadwood and saplings. We did this for a one-hectare plot at each of the sites, and three additional quarter hectare plots. The plots included a range of woodland types, from conifer plantations, to mature riparian broadleaf woodland, to Scots pine and Larch dominated woodland, to remnants of temperate rainforest. The data will be useful for verifying woodland assessment from satellite images, and for comparison with canopy height models generated from Lidar data.
In this video Annabel and Claudia explain our approach to the woodland surveys:
Grassland surveys
Left: Claudia recording species during a grassland survey
Grasslands can range from heavily grazed short cropped uniform fields to species rich taller swards. The ground-truth data we collected covered a mix of grassland types and condition levels. A total of 31 fields were surveyed across the three sites.
In this video Annabel explains our approach to the grassland surveys:
Peatland surveys
Left: A peat depth probe
The aim of the peatland surveys was to provide data to improve the use of remote sensing to detect where peat is, the condition of peatland and to monitor restoration. A total of seven sites were surveyed across the three estates, including open peatland, forest-to-bog restoration, and areas of shallow peat with deeper pockets. At each point on the sampling grid we recorded the depth of the peat, the peatland condition (for example near-natural, modified, drained or actively eroding), and the dominant vegetation (for example Cotton grass, heather or Sphagnum).
In this video, Claudia explains how we approach our peatland surveys
Saltmarsh survey
Left: A transect line for a saltmarsh survey
Saltmarshes are a vital intertidal wetland habitat which have importance with regards to both biodiversity and carbon storage. The NVC survey identified the areas of saltmarsh on the Tayvallich estate, but to understand the variability within the saltmarsh and condition factors such as grazing we carried out additional transect and quadrat surveys at 5 sites.
Left: Glasswort, a pioneer saltmarsh species
The transects ran from the low tide point to the highest point of saltmarsh habitat. They were divided into 5m intervals and a representative quadrat surveyed for each interval. We recorded the saltmarsh zone (e.g. mudflat/boundary, pioneer, low-mid marsh, mid-upper or terrestrial transition), each of the species present, and indicators of condition such as percentage of bare ground and grazing intensity. These records support the development of AI solutions that can not only identify saltmarsh, but also better represent its state.
We are planning to publish our research reviewing sources of error in ecological surveys in a journal later in the year.
The survey data now been fed into the TAiM website and portal, developed by EOLAS Insight Ltd, for all the information on the project see https://www.trustable-mapping.xyz/.
