Researcher · University of Alicante
I study how spatiotemporal AI systems can make dependable forecasts when sensors, data and deployment conditions are imperfect.
Research perspective
Marc Semper Lloret publishes academically as Marc Semper. He is an associate lecturer and researcher in the Department of Computer Science and Artificial Intelligence at the University of Alicante, and a member of the Network Data Analysis and Visualisation research group (ANVIDA).
His work connects graph learning, environmental observation and robust evaluation. The aim is not only to improve predictive accuracy, but to understand whether the selected model remains a defensible choice under realistic data and deployment uncertainty.
Read the research overviewResearch areas
Four connected lines of work centred on environmental systems, imperfect observations and accountable model choice.
Evaluation and selection methods for forecasting systems that must remain dependable under missing data, sensor failures, spatial misalignment and changing deployment conditions.
Graph neural networks that capture relationships across distributed environmental sensors, cities and global observational grids.
The effect of missingness, observational uncertainty, aggregation and measurement inconsistencies on learned models and the decisions built on them.
AI methods for air quality, atmospheric aerosols, greenhouse-gas concentrations, urban noise and other environmental phenomena.
Selected work
Journal of Computational Science, article 102988, 2026
Global atmospheric reanalyses like ERA5 exhibit persistent regional biases over complex terrain that can distort downstream hydrological, energy, and risk models. Combining local weather station networks (AVAMET) with spatial neighborhood graphs via GraphSAGE and Jumping Knowledge enables simultaneous correction across multiple meteorological variables. Rigorous chronological validation shows up to an 81% error reduction in pressure, humidity, temperature, and wind compared to classical baselines, while revealing clear operational boundaries for localized extreme precipitation.
Stochastic Environmental Research and Risk Assessment, vol. 40, no. 8, article 202, 2026
Correcting sub-daily satellite precipitation extremes (GPM IMERG V07) with machine learning becomes challenging when the ground observational reference itself carries spatial and density uncertainties. Benchmarking tabular models (LightGBM) and spatial deep learning (CNN) across strict temporal, spatial, and event-based holdouts demonstrates that continuous post-processing reliably reduces bulk error, but deterministic recovery of heavy rainfall tails hits fundamental limits. Direct probabilistic exceedance modeling provides the most dependable operational ranking for early warning and hazard screening.
Atmospheric Research, article 109244, 2026
Operational satellite precipitation products often struggle during Mediterranean flash floods due to rapid convective dynamics and steep orography. Auditing half-hourly GPM IMERG estimates against high-density rain-gauge networks quantifies pixel-scale spatial displacement, severe peak attenuation, and gauge co-availability limits, defining empirical boundaries for satellite-driven hydrological and flood modeling.
Knowledge-Based Systems, article 116673, 2026
Selecting the best forecasting model from a single clean validation set often leads to brittle deployment decisions when real-world evaluation data is missing, displaced, or degraded. Re-evaluating fixed predictions across declared reference perturbations, quantifying selection instability relative to training-seed variance, and balancing performance budgets transforms fragile model selection into an auditable decision-support system.
Mathematics, vol. 14, no. 5, article 908, 2026
Graph forecasting architectures typically train individual node embeddings in isolation for every separate dataset. Extracting subgraphs across multiple related spatiotemporal networks and transferring the learned representations enables cross-domain knowledge reuse, improving prediction accuracy on target tasks while providing explainable insights into the shared node space.
International Journal of Environmental Science and Technology, vol. 23, no. 1, article 69, 2026
Forecasting global atmospheric aerosols requires capturing both worldwide spatial transport mechanisms and multi-scale temporal dynamics. Combining CAMS aerosol records with ERA5 meteorological forcing in a hybrid architecture (MultiscaleTCNGraphSAGE) outperforms state-of-the-art transformer and graph baselines, yielding superior representation of extreme aerosol episodes.
Applied Sciences, vol. 15, no. 10, article 5576, 2025
Urban acoustic pollution exhibits strong cyclic temporal patterns driven by human activity alongside spatial correlations governed by street topology and urban morphology. Integrating 1D temporal convolutions, recurrent LSTM memory, and graph transformer layers (CNN1D+LSTM+TransformerConv) delivers highly accurate predictive mapping across Madrid's urban sensor grid.
Journal of Environmental Management, vol. 371, article 122922, 2024
Projecting global carbon dioxide and methane concentrations up to six months ahead demands architectures capable of synthesizing satellite observations with dynamic and static environmental covariates. Formulating the planetary observation system as an explicit spatiotemporal graph neural network provides superior predictive accuracy over purely temporal methods, underscoring the importance of global spatial connectivity in greenhouse gas dynamics.
Doctoral thesis
Spatiotemporal modelling with neural networks for forecasting environmental phenomena. Defended at the University of Alicante on 11 December 2025.
Explore the thesisAcademic identity