University of Bath Civil Engineering PhD Studentship: Machine Learning for Early Warning of Debris-Screen Blockage in Urban Rivers

University of Bath Civil Engineering PhD Studentship: Machine Learning for Early Warning of Debris-Screen Blockage in Urban Rivers

Quick Decision

This Bath Civil Engineering PhD asks whether a screen can be flagged before debris accumulation becomes a blockage, rather than merely classified once a camera shows it. Its distinctive combination is computer vision, rainfall and catchment information, probabilistic forecasting and hydraulic consequences. Thomas Kjeldsen's recent work supplies a credible detection baseline, while the unresolved step to a useful forecast is substantial. The advertisement explicitly requires both interest and experience in applying machine learning to technical problems; a relevant master's is advantageous rather than mandatory.

This is a concrete water-risk research question with international URSA competition. The topic is attractive if the available data support a genuine forecast, and an individual award still depends on selection. Apply through Civil Engineering by 29 November 2026; the October 2026 scheme terms specify 23:59 GMT. The expected funded start is 4 October 2027. First clarify forecast lead time, access to labelled events and the personal award's funded months. The current URSA terms and Civil programme are the relevant official routes.

Research Problem and Doctoral Work

Screens protect culverts from large debris but can themselves obstruct flow. The proposed scientific advance has three parts: recognising screen conditions across sites with little local training data; combining images with observed weather, rainfall forecasts and catchment descriptors to estimate future blockage probabilities; and coupling those probabilities to hydraulic models to examine downstream flood implications. Antecedent dry periods and rainfall intensity are proposed explanatory controls, not already established causal findings from this project.

A useful thesis would separate present-state classification from a forecast with a stated horizon and event definition. Randomly dividing photographs can test image recognition without testing performance on a later storm, another season or a previously unseen catchment. This is an assessment implication of the papers, not evidence that a particular dataset is contaminated. Probability calibration, missed dangerous events, unnecessary inspections and uncertain hydraulic consequences matter alongside average classification accuracy.

The advertisement promises analysis of large CCTV archives, meteorological data and real-world cases, but does not identify the sites, licences, synchronised maintenance logs, computational allocation or minimum forecast horizon. No public evidence establishes that this new workflow already reduces flood losses or staff exposure. The Environment Agency's culvert and screen research summary emphasises whole-life system assessment, including a presumption against new screens where alternatives exist. This is a case for evaluating existing infrastructure carefully, not assuming more cameras or more screens are always beneficial.

School, Programme and Field Context

The degree route is University of Bath Civil Engineering PhD, in Architecture and Civil Engineering within Engineering and Design. Bath is the degree-awarding university; URSA is a studentship mechanism rather than an advertised salaried employment contract. The programme normally takes four years full time, with a two-year minimum and normal four-year maximum. Confirmation includes a report and oral assessment around the first year, followed eventually by a thesis and viva. Registration duration and funded duration must be assessed separately.

Subject relevance comes from Kjeldsen's flood hydrology and uncertainty work and the Water Innovation and Research Centre, established in collaboration with Wessex Water. This demonstrates a water-research community, not a guaranteed placement, partner dataset or joint degree. The QS provider lists Bath at 125 in 2027, compared with 132 in 2026; these are overall rankings, not a hydrology or lab ranking. The university's doctoral research information reports 92% of its REF 2021 research as world-leading or internationally excellent, an institution-wide measure.

Doctoral professional development provides research and transferable-skills support. The personal curriculum in image analysis, hydrometeorology, probabilistic evaluation and hydraulics remains to be agreed. Applicants changing fields should establish who will teach the missing component and whether the first-year scope is achievable with the available data.

Supervisor and Lab Assessment

The October 2026 institutional profile identifies Professor Thomas Kjeldsen, accepting PhD students, with research in extreme events, hydrological modelling, machine learning and uncertainty. It gives T.R.Kjeldsen@bath.ac.uk. The advertisement header says Dr while its body says Prof; the current profile supports Professor. No additional formal supervisor or meeting frequency is publicly allocated here. Coauthorship and centre membership do not make every colleague part of this student's supervisory team.

Recent Important Publications

Smith and colleagues, 2025, CCTV image-based classification of blocked trash screens. This team paper, published online in 2024 and assigned to the 2025 journal issue, uses one Tongwynlais camera. Quality screening reduced 755 photographs to 577. Logistic-regression experiments with different class-balancing strategies reported 88% accuracy for the augmented case. The practical labels reflect a cleaning judgement. Selected primary methods, results, discussion and conclusions were read. Image exclusions and one-camera evaluation limit transfer claims; classification is not a verified future-event forecast. Some precision/recall wording and summary values are internally inconsistent, so those figures are not treated as reconciled operational benchmarks. See the published article.

Rowlatt and colleagues, 2026, Operational uncertainty in machine-learning based debris block detection. The same Cardiff screen and an overlapping period supplied 1,089 images, including visually uncertain cases. A hundred random 80/20 resamples compared logistic regression and neural models. Median accuracy for logistic regression was 78–79%, broadly comparable with the tested vision transformers; headline peaks above 80% are not routine warning reliability. Selected primary methods, results, discussion and conclusions show a trade-off between missing high-risk images and generating extra inspections. This extends uncertainty analysis, not independent multi-site replication or a rainfall-driven forecasting trial. Data and code are available on reasonable request. See the published study.

Vandaele, Dance and Ojha, 2024, Deep learning for automated trash screen blockage detection using cameras. This is an external Reading-team benchmark, not a Kjeldsen coauthored paper. Its publisher material describes 80,452 images from 54 cameras and compares binary, Siamese and anomaly approaches. The primary indexed methods and conclusions were read, not the complete PDF. The authors' repository identifies four held-out cameras, offers inference code rather than a training pipeline, and lists data and weights. Few labelled reference images used for adaptation are not zero-label universal transfer. The article gives a relevant comparison for the new project's cross-site ambitions; dataset access, licence and independent reproduction were not verified.

Together these papers support a real methodological starting point. They do not establish a deployable end-to-end forecast. Different image totals, sites and class definitions must remain separate; neither paper count nor repeated Cardiff samples supplies three independent validation populations. Exact permissions for operational archives, human labelling support and hydraulic models should be confirmed. Historical grants and computing acknowledgements in papers are not personal 2027 studentship resources.

Funding, Application, and Living Cost

The competition considers Home and Overseas applicants and can cover tuition, maintenance and up to £1,000 annual training support. Overseas awards are capped, with no percentage stated for this call. Training support is restricted research development money, not additional living income. The October 2026 URSA page gives Engineering and Design a general maximum of 3.5 years; it does not promise this individual 42 months. The advertised £21,805 is the 2026/27 rate, corroborated by UKRI guidance; the personal 2027/28 rate, award term and completion-period coverage remain unknown.

The 2027/28 Engineering and Design fee table lists £31,350 annually for full-time Overseas PhDs and allows continuing annual increases of up to 8%. Covered fees should not be counted twice in a budget, but any period beyond the award needs written clarification. The 2025/26 postgraduate budget is a dated reference: its published total is £405 weekly or £21,060 annually. Its halls-labelled total includes an off-campus utilities item, so £368 is an analyst's non-duplicated arithmetic scenario, not a replacement university total. Listed rent benchmarks—£230 weekly on a 51-week contract or £750 monthly off campus before bills—are not 2027 offers. Deposit, moving costs and first-payment timing still matter. The doctoral FAQ is also explicitly scoped to 2026/27.

International awardees must pay visa and Immigration Health Surcharge costs personally. The October 2026 Student visa guidance states a £558 application charge; the student IHS rate is £776 per year, with the upfront total depending on permission length and part-year rules. Future charges may change. These figures do not establish a reliable 2027 disposable-income surplus.

Applicant Fit and Preparation

Required Conditions and China Route

The exact project requires a held or forthcoming First/2:1 honours degree or international equivalent, and interest plus experience applying machine learning to technical problems. A master's is advantageous. Generic programme alternatives do not erase project-specific conditions. A credible application should show an individual contribution to a technical dataset or model, how it was evaluated and what failed; there is no advertised mandatory software language or named neural architecture.

The University encourages applications from under-represented groups. Applicants may explain circumstances that affected educational attainment in the application form; the notice recommends a short paragraph at the end of the personal statement. If an applicant states on the application form that they have a disability, the Disability Service will contact them to discuss their needs.

The Civil programme's international-equivalence route, Overseas URSA eligibility and the Student visa route provide positive grounds for Chinese applicants to be considered. Admission and funding remain separate decisions. The China entry information has an evidence gap: indexed PGR wording refers to a first-class bachelor's with research experience or a research-containing master's, while the current directly available country body does not reproduce a definitive China doctoral threshold. Do not reuse taught-master's percentages or assume a master's automatically overrides an undergraduate requirement. Send the actual institution, degree and transcript to admissions.

English, Immigration and Application Evidence

Civil normally requires IELTS 6.5 overall with every component at least 6.0. The research English policy distinguishes TOEFL tests before 21 January 2026—90 overall, 21 each—from tests on or after that date—4.5 overall, 4 each. Accepted tests normally need to be within 24 months of entry. A qualifying English-medium degree can support an individually confirmed exemption; Chinese medium-of-instruction evidence is not an automatic waiver.

The exact advertisement is stricter about timing: non-UK applicants who are not currently studying in the UK must meet the English requirement before the application deadline, with no exceptions. A routine conditional English offer must not be assumed to rescue an incomplete funding application. Civil is listed as ATAS-covered; China is not nationality-exempt under the ATAS rules. The assigned course code and clearance are individual matters, not a nationality ban or an approval guarantee.

Use Bath's Civil Engineering application. Select University of Bath URSA in funding, enter the exact title and lead supervisor, and use the designated second-project fields if applicable; this competition permits at most two projects. The application guide requires supporting qualifications, translations where needed, a personal statement and two references, including an academic reference from the institution where the applicant studied most recently; intended PhD supervisors must not be referees. Its at-least-1,000-word proposal rule is for an applicant's own idea, not an automatic extra rule for every advertised project. Applicants must write their own application; permitted AI assistance does not authorise generated application content.

Next Actions and Questions to Verify

Before applying: ask Kjeldsen which forecast horizon and blockage event define the initial thesis, whether the image and weather archives are already licensed and linked to maintenance records, and how prior ML experience will be assessed. Confirm individual degree equivalence, accepted English evidence and the advertised-project proposal requirement with admissions. Submit the correct URSA fields before the deadline.

At interview: establish the first-year baseline, independent event/site evaluation and false-alarm trade-off; ask who supervises hydraulics, statistics and vision, and what computing and labelling support is committed. Existing camera predictions should not be presented as validated intervention benefits.

Before accepting: obtain the personal stipend rate, funded months, fee/completion coverage and first-payment schedule in writing, alongside ATAS/CAS timing and a realistic housing budget. Any alternative entry date also needs confirmation that funding survives the change.

The transferable route is flood-risk analytics, environmental computer vision, uncertainty assessment and infrastructure asset management. The Environment Agency's 2025 guidance recognises technology-assisted monitoring within resource-constrained maintenance. That supports professional relevance in authorities and consultancies, not an established placement, salary or employment guarantee.

Evidence and Sources

Recruitment and policy information is from October 2026. The original-paper discussion retains the reading limits stated above; current vacancy availability, archive licences and independent reproduction have not been established.

Project, supervision and original research

• Thomas Kjeldsen current profile
• Smith et al 2025 primary article
• Rowlatt et al 2026 primary article
• Vandaele et al 2024 primary article
• Authors inference code repository
• Bath WIRC institutional centre
• EA culvert and screen manual summary
• EA 2025 asset inspection guidance

School, training and application

• Civil Engineering PhD programme
• PGR English language requirements
• International and China entry information
• Doctoral professional development
• QS Bath ranking by edition
• Bath postgraduate research and REF context
• ATAS immigration rules
• Bath doctoral application guide

Funding and living costs

• Current URSA funding competition
• 2027/28 Engineering and Design tuition
• 2025/26 postgraduate budget
• Doctoral application FAQ
• UKRI training grant guidance
• Student visa official guidance
• Student immigration health surcharge