Our projects connect a clearly defined biological or environmental problem to an appropriate measurement strategy, defensible analysis and a practical decision. Public summaries are intentionally concise where work is unpublished.
Repeated UAV observations, radiative-transfer modelling and knowledge-guided machine learning are being combined to infer crop physiological state at breeding-population scale. The work links automated plot phenotyping with explicit crop-growth pathways so that early selection can be based on interpretable routes to performance, not image prediction alone.
Leaf spectroscopy, UAV multispectral sensing, radiative-transfer models and temporal machine learning are being integrated to distinguish emerging disease effects from other sources of crop variation. Dynamic crop modelling then connects those observations to production forecasts and actionable surveillance.
Programmatically controlled photogrammetry is used to turn processing choices, early diagnostics and geometric quality measures into an experimental dataset. Machine learning investigates whether costly reconstruction failures can be predicted before full point-cloud and mesh generation, with immersive environments used as a demanding fidelity test.
Spatial data science and network analysis are being used to study how carbon-accounting tools, advice and trusted intermediary support move through agricultural knowledge networks. The project combines mapping, graph-based modelling and uncertainty-aware statistical and machine-learning workflows to investigate reach, overlap and opportunities for future collaboration.
A crop-growth graph retains explicit light capture, radiation conversion, biomass formation and partitioning while neural parameter-generators adapt the mechanisms to local observations. Scientific auditing uses independent physiological anchors, structural tests and counterfactual response analysis alongside prediction.
Longitudinal phenotypes are used to construct a stable, target-independent reference geometry before new observations are introduced. The programme tests cross-environment persistence, placement of unseen genotypes, multiple trait annotations and relocation through alternative sensors without refitting the biological map.
A differentiable PINN architecture embeds established empirical and ensemble formulations as structured prior components, then uses constrained neural pathways to learn residual nonlinear behaviour. Auditing through ablation, controlled systems and independent validation tests whether flexibility is gained without erasing scientific meaning; application-specific details remain undisclosed while the first manuscript is in preparation.
Controlled radar acquisition, depth-resolved signal processing and mechanistic feature engineering are used to characterise how subsurface stone burden transforms the electromagnetic wavefield. Machine learning converts those signals into uncertainty-aware operational evidence for spatially selective intervention.
National flood assessments are harmonised with small-area evidence on social conditions, accessibility, warning and protection. Spatial statistical models distinguish river, coastal and surface-water processes while accounting for the geographical dependence that undermines naive rural-urban comparisons.
Greenbox combines autonomous solar-powered recorders, mobile connectivity, cloud infrastructure, AI-assisted species detection and reproducible ecological analytics. Event collapsing, threshold sensitivity, historical context and expert review safeguards turn continuous high-throughput sensing into cautious evidence about community structure, phenology and conservation interest.
