2025-C1 Digital Ecological Landscapes

Overview

This research aims to automate the production of high-resolution habitat maps for diversity monitoring across UK landscapes. Accurate, scale dependant, information on habit location and condition is essential for understanding species distribution and movement, and to effectively target resources for nature recovery. This is urgent as there has been a significant loss of biodiversity within the UK, with a 60% decline in the abundance of UK priority species since 1970. 

Recent advances in image-based AI allow us to evaluate the extent and distribution of habitats faster, more efficiently and with higher accuracy (van de Plas et al 2023, see figure 1). However, the variation in landcover caused by climate, seasonality and management requires a corpus of examples to train robust models. When combined with the limitations of image data collection (variation in timing, resolution, quality) and the semantics of habitat types, the lack of a high quality, representative datasets are a major limitation to the scaling of high-resolution predictions to the whole of the terrestrial UK (or globally). 

This project aims to create an predictive model for UK landscapes through: 1) investigating the scaling of small, project specific datasets for Machine Learning to large scale datasets suitable for training models with high generalisation, 2) evaluating the training performance of fully supervised, semi- and unsupervised training methods, 3) evaluating the effect of image resolution and timing on model predictions, and 4) critically evaluating model outputs for monitoring biodiversity. The project is supported by the Peak District National Park Authority, with image data supplied by Aerial Photography GB and Airbus Defence and Space. 

Figure 1: Landscape information at high resolution from a landcover segmentation model (see van de Plas et al 2023) using aerial photography at 12.5 cm resolution. Background image Aerial Photography Great Britain.