PhD: Quantifying uncertainty and spatial dependence in tropical forest forecasts to guide climate change mitigation

PhD: Quantifying uncertainty and spatial dependence in tropical forest forecasts to guide climate change mitigation

Research Question and Application Judgment

How uncertain is a forest-loss forecast at one location, and how do related outcomes across locations change a conservation decision? This Exeter doctorate connects those statistical questions to the allocation of scarce conservation funding. The fit may be weaker for applicants seeking a purely computational doctorate with little engagement in environmental economics, causal inference or policy constraints.

The project information identifies 2027, University of Exeter, Land Environment Economics & Policy Institute (LEEP), Ben Balmford as primary supervisor, and Ben Groom and Daniel Williamson as academic supervisors. Carbon at Risk and Bali Lee are named project partners; the CASE field is blank. Partner names do not establish a CASE supplement, guaranteed placement or access to commercially usable data. Exact native project

The official schedule lists 16 November 2026 as the opening date and 13 January 2027 at 13:00 GMT as the deadline. The opening date is not the cutoff. An official notice says not to submit before the system opens; the detailed guidance still contains 2026 material, so confirm the 2027 forms and codes before applying. Official application process

International applicants, including Chinese nationals in principle, can compete for the full-fee scholarship. Their places are limited and highly competitive. Individual academic eligibility remains conditional, particularly for a Chinese bachelor-only or nonstandard Master's background: the CDT's published requirements and Exeter's generic host requirements are different and must be reconciled in writing.

Scientific work and the important methodological distinction

The advertised baseline uses Random Survival Forests and satellite observations to forecast annual deforestation and regrowth probabilities. The proposed doctorate tackles uncertainty around a particular pixel's prediction and the joint spatial distribution of forest outcomes. These are separate problems: a model's out-of-sample error describes aggregate predictive performance, whereas a decision may depend on uncertainty at a specific site and correlated outcomes across several sites.

The advertisement proposes BART, deep kernel learning and variational autoencoders as possible foundations, and neural-network methods for spatial dependence and spillovers. These are candidate research directions, not a promised final architecture. The intended downstream product is a portfolio decision tool choosing interventions with high expected additionality while diversifying correlated risk and limiting displacement. Its usefulness will depend on how the statistical uncertainty is represented and which policy assumptions enter the objective.

A key distinction is between correlated forest outcomes and causal displacement. A shared economic or weather shock may correlate losses at neighbouring sites. An intervention can also move deforestation pressure elsewhere. Residual spatial correlation alone does not establish that protection caused that movement. Likewise, accurately forecasting business-as-usual forest loss does not establish how much a particular intervention would prevent. The thesis should explain where forecast uncertainty ends and assumptions about treatment effects begin.

A sensible preparation exercise would compare a simple calibrated baseline with a spatial extension on held-out years and regions, then show how the joint forecast changes a conservation allocation. This is analyst advice, not an additional admission requirement. Scientific success should be judged using both predictive diagnostics and the value or robustness of the decisions, with sensitivity to intervention-effect assumptions.

The project description does not settle dataset geography, licences, observation errors, access dates or partner publication rights. The linked BIOADD programme describes dynamic-landscape additionality and policy tools, which establishes an applied research context rather than guaranteed availability of all its underlying data. BIOADD programme

Three Original Studies and Their Limits

Evaluate the allocation rule, not only prediction

Mancini and colleagues, One Earth, 2024. The original compares five spatial targeting rules for biodiversity offsets in England using biodiversity, agricultural opportunity-cost and recreational-access models. Selected results show that locally constrained offsets perform poorly and that alternative rules expose biodiversity–equity trade-offs. Analysis uses a 400-hectare grid because confidential agricultural records limit finer costing; some inputs cannot be shared. This supports evaluating the decision rule, not merely forecast accuracy, but it does not validate tropical forest prediction or identify tropical leakage. Balmford and Groom are coauthors. The contribution statement credits Mancini and Collins with analyses, other named authors with model development, and all authors with collaborative writing; it does not assign Balmford or Groom primary coding responsibility. Published paper PDF

Temporary storage and carbon value

Groom and Venmans, Nature, 2023, The social value of offsets. The selected accepted manuscript derives welfare value from the timing of avoided and later released emissions, incorporating failure and non-additionality hazards. Its numerical examples show that temporary storage can have value without being equivalent one-for-one to permanent removal. Equivalence depends on economic, climate and risk assumptions; precise context-specific additionality estimates remain difficult. This supplies a valuation framework for a conservation portfolio, not an empirical guarantee that any advertised offset is additional. The accepted manuscript attributes 25% of theory to Groom and 75% to Venmans, with equal contributions on its other reported aspects. Its wording is version-specific, so it is not substituted for the final publisher's contribution statement. Author accepted manuscript

Accuracy is not uncertainty calibration

Yazdi, Bingham and Williamson, 2024 original preprint. This study modifies deep Gaussian-process emulation with a layer-smoothness parameter and stochastic variational inference for large, nonstationary simulations. In the COMPAS astrophysics example, the three-layer model explains 95.76% of response variation, but its nominal 95% coverage is 90.4%, compared with 91.7% for two layers. This illustrates why predictive accuracy and uncertainty calibration should be checked separately. Errors around zero/nonzero transitions remain important. It is an author preprint, not confirmed peer-reviewed forest evidence, and transferring deterministic-simulator emulation to observational forest data requires additional work. Williamson is a named coauthor with Exeter Mathematics affiliation; individual coding, analysis or supervision duties are not reported in the selected text. Dated author preprint, version 1

Together, these studies provide relevant context on decision evaluation, carbon valuation and uncertainty. They do not establish that the proposed spatial forest model already works.

University, discipline, degree and training

The awarding institution would be the University of Exeter in the United Kingdom; UNRISK is a multi-university training centre managed by Leeds, not the degree-awarding university for this Exeter project. Its exact native institutional setting is LEEP. Balmford's native profile places LEEP within the Economics Department. Confirm the precise Exeter registration programme and its progression rules; the generic Economics MPhil/PhD description is context, not proof of this project's administrative code. Balmford's university profile

Exeter reports 136th in QS 2027 on its current China information page. That is an overall university ranking. University ranking context The generic Economics page reports more than 91% of outputs at REF 2021's 3-star/4-star level; it is dated subject evidence rather than a score for this project or its supervisors. Interdisciplinary fit and access to methods support matter more directly to this choice. Economics doctoral context

UNRISK starts doctoral research immediately alongside cohort training, concentrated early in the doctorate. Its programme includes residential workshops, peer learning and partner Challenge Weeks. Subject modules may be attended without assessment; cohort events do not themselves count towards the PhD award. Students register at a host university and have at least two academic supervisors. The programme expects a three-month placement within the funded period. Obtain the actual 2027 calendar, placement arrangements and Exeter training plan rather than assuming every listed opportunity becomes an individually guaranteed course or partner experience. UNRISK training structure

Supervisors, laboratory resources and future uses

Balmford is a Lecturer/Assistant Professor in Economics. His native profile explicitly connects machine learning with quasi-experimental methods for forest dynamics, additionality and leakage, alongside environmental mechanism design. His personal research list includes ongoing dynamic-forest additionality work. The named primary supervisor therefore has a direct scientific connection to the advertised question, rather than being inferred from a nearby department's publications. Personal research list

Groom and Williamson appear in LEEP's leadership and research environment. The institute combines economics with natural, social, physical and data sciences and describes policy-facing tools and external partnerships. This is relevant to a student wanting methods translated into decisions. Institute impact figures and partner counts do not provide completion rates, supervision frequency or individual placement outcomes. LEEP environment

Before committing, establish who helps day to day with statistical theory, geospatial computing and economic identification. Three named supervisors do not guarantee equal availability or a particular meeting schedule. Ask whether the student will inherit a maintained codebase, which computing resources are allocated and how methodological papers and partner-facing software are balanced.

Funding duration, international fees and living costs

The published scholarship covers full tuition fees, including the international difference, and UKRI-standard stipend for three years nine months. The expected three-month placement is included within that period, not added afterwards. Do not market this as a four-year stipend without further award evidence. The £6,000 individual Research Training and Support Grant and £5,000 per-student cohort-training allocation are programme research/training budgets, not annual supplements to personal living money. Flexible-fund access is an opportunity rather than a personal cash promise. Specific funding terms

The FAQ expressly applies to 2027 entry while warning that historical timings can change. It confirms fee-gap coverage and excludes relocation-related costs such as accommodation, travel, visa and health surcharge. Coverage depends on winning the UNRISK scholarship; a self-funded place or ordinary host admission is a different financial route. Exact award wording, funding end date and any writing-up period beyond the funded term should be obtained before acceptance. Funding FAQ

Current UKRI guidance lists £21,805 for 2026–27, around £1,817 monthly, and a Home-fee funding level of £5,238. It does not yet give the 2027–28 stipend in the table read. Annual rates can change; no flat multi-year total or individual net income is inferred. UKRI says international cohort limits vary by council. UNRISK confirms that international awards are restricted, but the available project terms do not state a numerical quota; no unrelated council percentage is applied to this project or converted into applicant odds. Current UKRI rates and limits

Exeter's budget, based on April 2026 local averages, estimates £1,300–£2,000 monthly for basic single-student living expenses, or £15,600–£24,000 over twelve months. Compared with the same-period £21,805 stipend, the arithmetic leaves £6,205 at the low end or a £2,195 deficit at the high end, before exceptional expenses. This is a dated illustration, not a 2027 affordability guarantee. Private rooms are quoted at £135–£250 weekly excluding bills; inspect the real contract, length and deposits. Obtain a housing plan for the October start and a reserve for arrival costs and any unfunded writing-up. University living-cost estimate

The current outside-London visa maintenance figure is £1,171 monthly for up to nine months, or £10,539, plus applicable unpaid course fees. That is documentary eligibility, not a twelve-month living-cost budget. Sponsorship evidence and China's differential documentary treatment may affect what initially needs to be submitted; the underlying financial condition remains, and evidence can be requested. The currently displayed Student visa fee is £558, but 2027 fees, health surcharge and charged permission duration require checking nearer the application. Visa money rules

Chinese-applicant conditions and preparation

No Chinese-nationality exclusion was found: the exact project belongs to a scheme explicitly open to international applicants and covering their full tuition if awarded. The CDT accepts a relevant 2:1 honours degree or non-UK equivalent and says a Master's is not universally required. It excludes people already awarded a PhD or currently registered for one; a previous unfinished doctorate may be considered.

However, Exeter's generic Economics page asks for a relevant taught Master's, and its China research-degree guidance normally expects a completed three-year Chinese Master's with both certificates. The current Chinese page's 75% undergraduate equivalence concerns taught postgraduate admission and must not be used as an automatic bachelor-only PhD permission. The right conclusion is conditional eligibility requiring the CDT and host to determine the applicable route and personal qualification, not automatic rejection or automatic acceptance. China qualification guidance, UKRI host responsibility for eligibility

There is also an English-profile difference. UNRISK's Exeter button links Profile A, with IELTS 6.0 overall and 6.0 in every component. Generic Economics points to Profile B2, with 6.5 overall and 6.0 in every component. The exact offer/registration route must resolve this; neither is assumed to be the individual's final condition. An English qualification need not be complete at initial application but must meet host requirements before October registration. Profile A, Profile B2

China is not an ATAS-exempt nationality. Applicability depends on the exact course subject code and CAS; neither “Economics” nor the presence of neural networks settles it. Confirm the code and any clearance requirement with Exeter, and allow for conditional admission and immigration lead times. ATAS rules

A convincing applicant should show quantitative programming and a carefully validated analysis, explaining model assumptions and why uncertainty affects a real decision. Spatially and temporally separated validation, calibration, reproducibility and sensitivity analysis are useful evidence. These are analyst preparation priorities, not newly imposed official prerequisites. Economics applicants should demonstrate computing competence; machine-learning applicants should explain how they will distinguish prediction from causal policy evaluation.

Career Relevance

Potential uses of the training include environmental economics research, geospatial modelling, climate-risk analytics and decision-support software. This is an inference from documented work and programme skills, not a hiring forecast or salary estimate. The placement could help test non-academic preferences, but it neither guarantees employment nor proves the named partners will host this particular student.

Application Stages and Decision Points

All UNRISK hosts use the Leeds online administrative application plus a UNRISK-specific assessment form. The published material permits up to two project choices and names NERC UNRISK CDT as the planned course. Its detailed questions and upload rules are still labelled legacy 2026 and must be checked after the announced update. The old body uses a structured statement rather than a conventional CV and does not require initial references; successful Exeter candidates later make a host application that may require references. Generic Economics proposal, supervisor-letter and deadline rules must not silently be substituted for the CDT's initial process.

The current 2027 timetable lists interview invitations 18–26 February, an interviewee open day 1 March, interviews 3–5 March, and offers 10–17 March. Programme resources refer to October registration, supporting October 2027 entry, but the exact start day and host registration conditions remain to confirm.

Decision area	Assessment	Evidence still needed
University and programme	Relevant Exeter LEEP setting and cross-university cohort	Exact degree registration and 2027 training/progression
Supervisors and resources	Named team connects conservation economics and uncertainty methods	Availability, computing allocation, inherited data/code
Research direction	Strong uncertainty-to-decision problem	Calibration, causal displacement and intervention assumptions
Funding and costs	Full international fees if awarded; 45-month stipend	2027 rate, housing, placement expenses and writing-up gap
Applicant fit	Quantitative international route exists	Personal degree equivalence and English profile
Risks and action	Prepare for 16 November opening	Updated forms, award competition, ATAS/CAS and partner rights

Ask for the precise Exeter degree route and qualification assessment; the applicable English profile; existing forest data and code access; partner licences and publication restrictions; measures distinguishing causal leakage from spatial correlation; supervision responsibilities; computing capacity; and the award's fee coverage, stipend, placement costs and funded end date.

Prepare a focused application only after clarifying the host qualification route and a credible data/methods plan. A listed project is not a funded offer. The strongest fit is an applicant who wants statistical methods to change real conservation decisions while keeping prediction, calibration and causal intervention assumptions distinct.

Evidence and Sources

The cited project, programme, paper, funding and cost sources have different dates and scopes. Application conditions, award terms, eligibility and immigration requirements should be confirmed with the relevant official body for the 2027 intake. The available evidence does not establish access to partner data or the effectiveness of any proposed intervention.