Earth Observation and deep learning: towards biodiversity enhancing, low carbon farming and land-use systems
Supervisors: Dr Nabi Omidvar, School of Computing and Leeds University Business School, University of Leeds, Prof Ed Mitchard, School of GeoSciences, University of Edinburgh, Prof Iain Clacher, Leeds University Business School, University of Leeds, Dr Gesa Reiss, Global Food and Environment Institute, University of Leeds. Industry partner: HD Wool Ltd (tbc)
Scientific background and motivation
Agriculture is one of the largest contributors to biodiversity loss. The biggest threat arise from clearing new areas for agriculture (causing direct losses, and indirect losses through habitat fragmentation), intensification of agriculture, and by releasing pollutants, including greenhouses gases (GHGs). Methods for biodiversity-friendly, low carbon farming systems exist, e.g., globally regenerative agriculture is promoted as a solution. However, transitioning conventional farming systems to regenerative ones as well as upscaling these systems is a major challenge. Nature-based Solutions (NbS) have been identified by IPCC as one of the most cost-effective climate change mitigations needed, yet NbS currently only receive 3% of global climate investment. The finance sector, whilst mostly concerned with landscape scale approaches, would be able to utilise regenerative agriculture as a testbed to find pathways to invest in nature and biodiversity.
Validating the transition from conventional to regenerative farming systems requires a shift from a focus on inventory-style management practices to a system based on measurable outcomes (e.g., soil health, biodiversity). Progress of this transition is slow due to lack of robust metrics and technology as well as to complex decision making to support new business models.
Aims and objectives of the PhD project
This project will aim to investigate novel ways of using Earth Observation (EO) data and existing derived EO products, combined with machine learning (ML), to enhance the cumbersome process of quantifying biodiversity, soil health, and ecosystem functions within the Ecological Outcome Verification (EOV) framework (2021) developed by the Savory Institute, US. This will assist in quantifying the current ecological impact of agriculture across specific study countries (e.g. a contrasting set such as the UK, Kenya, South Africa, Vietnam, Paraguay), and creating a spatial dataset that will give a baseline from which farmers could attract funding or better terms from financial organisations in order to transition from conventional to regenerative farming systems. This project seeks to use spectral, spatial, and temporal resolution capabilities of EO technology, and ML to replace or simplify the complex on-farm data collection currently required for EOV evaluation and verification. For instance, the power of deep learning combined EO datasets, e.g. vegetation indices such as NDVI from Sentinel-2 (10 m resolution, 5 day cadence), dynamic landcover data (e.g. Dynamic World, 10 m resolution), or Synthetic Aperture Radar from Sentinel-1 (10m resolution, 12 day cadence or better) to provide details on the biodiversity and ecological value of farms. Advanced multicriteria modelling techniques can be used to devise new metrics for EOV evaluation.
We will primarily use satellite images along with the data collected by farmers for EOV. This will allow us to create a labelled dataset of EO data tagged with EOV assessment results and train ML models such as Random Forests or more complex deep learning (DL) models such as convolutional neural networks (CNN) to investigate the efficacy of EO for EOV assessment. We also plan to collect geo-tagged audio signals, e.g., bird calls, of reference areas in Ecoregions and their associated EO imagery to build proxy metrics for biodiversity and soil quality. DL are prominent tools in representation learning, and we believe CNNs combined with sequence modelling techniques such as recurrent neural networks and transformers, can play a significant role in automating biodiversity and soil health assessments using EO data augmentation.
Project Partner – HD Wool Ltd (TBC)
 Benton, Tim G., et al. “Food system impacts on biodiversity loss. Three levers for food system transformation in support of nature.” Chatham House, London (2021).
 Breiman, Leo. “Random forests.” Machine learning 45.1 (2001): 5-32.
 Krizhevsky, Alex, et al. “ImageNet classification with deep convolutional neural networks.” Communications of the ACM 60.6 (2017): 84-90.
 Hochreiter, Sepp, and Jürgen Schmidhuber. “Long short-term memory.” Neural computation 9.8 (1997): 1735-1780.
 Vaswani, Ashish, et al. “Attention is all you need.” Advances in neural information processing systems 30 (2017).