Artificial Intelligence for Satellite Assessment in Grassland Ecosystems (AI-SAGE)
Dr Robert Atkinson, University of Strathclyde; Dr Lan Qie, University of Lincoln; Prof Christos Tachtatzis, University of Strathclyde; Lactalis
Research Aims
​
Satellite Earth Observation (EO) Data has emerged as a game-changer in accurately assessing plant biodiversity and ecosystem functions within agricultural landscapes, spanning both grassland and arable terrains. Near ground image acquisition from platforms such as drones, vehicle mounted devices or handheld camera phones, provide unparalleled spatial and/or spectral resolution, capturing extreme details, however, their frequent deployment for high temporal monitoring pose operational and cost challenges. Conversely, satellite data provides a frequent temporal overview—often revisiting the same spots multiple times a week, while recent satellite missions are delivering higher spatial resolutions, boasting Ground Sample Densities of tens of centimetres per pixel and can survey vast global expanses in one pass, facilitating broader and more encompassing studies of agricultural biodiversity. Furthermore, combining data from multiple satellite missions provides the ability to increase the spectral resolution effectively operating as a virtual multispectral instrument outperforming single satellite mission data source. The current challenges in harnessing EO data through Deep Neural Networks (DNNs) is not a scarcity of data, but the limited availability of detailed, pixel-level annotations.
​
The studentship aims to:
- Extract biodiversity indicators in grassland landscapes.
- Bridge this gap between large volumes of data and limited pixel level annotations, by synergising ground truth measurements derived from physical soil and grassland samples with the intricate details from near ground imagery and permit the training and validation of EO models alleviating the annotation shortage.
- Explore domain shift: EO trained models typically do not transfer between geographies and in the context of grassland species biodiversity as new species are introduced in the pasture this may result in out-of-domain model drift. Approaches such as internal learning with calibration/anchoring through physical sampling and near ground imagery will be investigated.
- Develop refined strategies and guidance on management strategies of the grasslands (inc. rotation, fertiliser application, etc.)
​
​