Background

Slowing tropical forest loss and enhancing regrowth are central to climate mitigation, but resources for protecting and restoring forests are limited. Both public conservation programmes and private carbon finance therefore need robust forecasts of where forests are most likely to be lost, or least likely to regrow, so interventions can target places where they may genuinely improve outcomes relative to business as usual. Recent advances in data make it possible to model forest-cover change at fine spatial and temporal scales. However, useful decisions require more than just accurate point forecasts: we need to quantify uncertainty and understand spatial correlation and displacement effects. Developing statistical and machine-learning methods that capture these uncertainties is essential for directing conservation finance effectively, restoring credible carbon markets, and designing portfolios of interventions robust to shifting economic and environmental conditions.

PhD project

This PhD will develop statistical and machine-learning methods to improve how forest conservation is targeted under uncertainty. Our existing work uses Random Survival Forests (RSFs) and satellite data to forecast annual probabilities of deforestation and regrowth, helping target interventions where they are most likely to deliver additional climate mitigation. However, two methodological challenges remain.
First, although RSFs produce probabilistic forecasts of forest change, they do not naturally quantify uncertainty around pixel-level predictions. Out-of-sample model error indicates overall predictive performance, but not how uncertain an individual forecast is. Second, the framework does not represent the joint spatial distribution of forest outcomes. Forecasts based on baseline characteristics avoid predicting intermediate drivers and propagating their errors forward, but treat neighbouring locations independently. This misses correlated shocks, and spatial spillovers, where protection in one location displaces deforestation pressure elsewhere.
The PhD will extend and adapt statistical and machine-learning methods to retain the flexibility and scalability of current approaches while providing calibrated, spatially explicit uncertainty quantification and addressing both forms of spatial dependence. Recent advances in BART, deep kernel learning and variational autoencoders provide promising foundations for maintaining predictive skill while characterising uncertainty. It will also develop neural-network-based methods to capture spatial dependencies and spillovers. These methodological advances will then be applied to conservation planning, culminating in a portfolio-based decision-support tool that identifies combinations of interventions that maximise expected additionality while diversifying correlated risks and limiting displacement, with applications to public conservation spending and private carbon finance.

Applicant Profile

We are particularly interested in students with a strong background in statistics, mathematics, computer science, economics, or a related quantitative discipline, who want to apply and develop these skills to address major environmental challenges. The project would suit applicants interested in statistical or machine-learning methodology, spatial data and uncertainty quantification, who have an appetite to translate methodological advances into practical tools for nature-based climate policy.

Other Information

https://dragonchair.org.uk/project/bioadd/
https://www.researchsquare.com/article/rs-7714530/v2