ERU GeoAI Research Group is an interdisciplinary research group that brings artificial intelligence and nature-inspired optimisation to problems in the Earth sciences and engineering. We work with satellite and hyperspectral imagery, climate and hydrological records, and geotechnical measurements. Accuracy matters to us, but so do three other things: our models should be interpretable, uncertainty-aware and fully reproducible.

The group builds on original metaheuristic optimisation algorithms that its members have contributed to the international literature. We do more than test these algorithms on benchmark functions. We build them into machine learning models, where they shape model structure, tune hyperparameters, search neural architectures and discover governing equations directly from data.

Our Vision: To become an internationally recognised centre of excellence in GeoAI, turning spatial data into trustworthy knowledge that informs field practice and decision-making.

Our Mission: To apply original optimisation methods to real problems in the Earth sciences and engineering; to test every method under rigorous statistical protocols; to practise open and reproducible science; and to train the next generation of researchers.

Research Themes

1. Nature-Inspired Optimisation and Automated Machine Learning

• Building metaheuristic search into ensemble learning and tree-based models
• Surrogate-assisted neural architecture search and hyperparameter optimisation
• Hybrid, multi-objective and parallel optimisation frameworks that combine Bayesian optimisation with evolutionary computation
• Support vector machines whose kernel functions are designed to suit each dataset

2. Interpretable AI and Symbolic Regression

• Symbolic regression methods that derive closed-form, human-readable engineering equations directly from data
• Hybrid models that pair equation-based learners with probabilistic methods, giving both interpretability and calibrated prediction intervals
• Explainable AI tools such as SHAP, used to give model behaviour a physically meaningful interpretation

3. Remote Sensing and Hyperspectral Image Analysis

• Land-cover and surface-water classification from hyperspectral and multispectral imagery
• Validation that resists spatial leakage: neighbouring pixels are strongly correlated, so random cross-validation inflates reported accuracy. We develop spatially disjoint validation protocols that measure how well a model generalises to new locations.
• Deep learning estimates of land surface temperature and other environmental variables, using cloud-based geospatial platforms such as Google Earth Engine

4. Climate, Hydrology and Environmental Risk

• Machine learning models for estimating and forecasting drought indices
• Spatio-temporal analysis of urban heat islands and climatic variables
• Feature attribution and uncertainty analysis for environmental prediction models