IAP-26-043

Geospatial Foundation Models in Geomorphology: New Insights into Controls on Landscape Denudation

Understanding the processes that control landscape denudation – the wearing down, stripping, and lowering of the Earth’s surface – remains a fundamental challenge in geomorphology. Researchers have used a range of topographic and environmental variables to explain spatial variations in observed denudation rates, yet these variables account for only part of the observed variability. This limitation suggests that important landscape characteristics and environmental interactions remain unrepresented in conventional approaches.

Recent advances in geospatial foundation models, such as AlphaEarth, offer a new way to represent these characteristics. Trained on diverse Earth observation data, these models generate global representations of the Earth’s surface that may capture spatiotemporal landscape patterns and environmental interactions. As such, these representations could improve our ability to explain spatial variations in denudation rates. However, whether they offer better predictive performance, generalise across contrasting landscapes, and provide new insights into the processes governing denudation remains largely unknown.

This PhD project will evaluate the potential of geospatial foundation models to improve both the prediction and understanding of landscape denudation rates. By integrating published global inventories of denudation rates with foundation-model representations and established climatic, geological, and topographic predictors, the project will compare foundation-model-based approaches with conventional methods and investigate what additional information these models provide about landscape denudation.

The project will address three overarching objectives:
– Determine whether geospatial foundation models explain spatial variations in observed denudation rates better than conventional approaches,
– Assess whether combining foundation-model representations with established climatic, geological, and topographic predictors provides additional explanatory power.
– Investigate whether relationships learned from foundation-model representations remain robust across geographically distinct regions with contrasting climatic, geological, and topographic settings.

The project will integrate published global inventories of landscape denudation rates with geospatial foundation-model representations, such as AlphaEarth, and conventional climatic, geological, and topographic metrics. These spatial datasets will be compiled and processed using Python libraries such as Google Earth Engine, GeoShapley, and TopoToolbox. Machine-learning algorithms, such as Random Forest, gradient-boosted trees (XGBoost), and neural networks, will be trained to predict denudation rates using three predictor sets: conventional environmental variables, foundation-model representations, and their combination. Comparing predictive performance will establish whether foundation models capture additional information about denudation patterns.

Model performance will be evaluated using cross-validation, including spatial block cross-validation, to assess predictive accuracy and generalisation across geographically distinct regions. Further tests across contrasting climatic, geological, and topographic settings will identify the environmental conditions under which model predictions are most reliable.

Finally, model interpretation techniques, including SHapley Additive exPlanations (SHAP) and permutation feature importance, will help identify which predictors contribute most to model performance. Ablation experiments, which systematically remove or combine groups of predictors, will quantify the additional explanatory power of foundation-model representations. By examining how these contributions vary across regions and environmental settings, the project will assess whether foundation models capture landscape characteristics and interactions that conventional predictors overlook, providing new insights into the spatial variability of landscape denudation.

Year 1

Review literature; compile global denudation, climatic, geological, and topographic datasets, acquire foundation-model representations, and develop the geospatial data-processing workflow. Attend relevant training workshops.

Year 2

Develop and train machine-learning models; compare conventional predictors, foundation-model representations, and their combination. Attend national conference and present findings.

Year 3

Evaluate model generalisation across geographical regions, interpret model predictions, and investigate the additional information captured by foundation models.

Year 3.5

Synthesise findings, prepare research publications, and complete and submit the PhD thesis.

This interdisciplinary project will enable the student to develop expertise at the intersection of Earth science, geospatial data science, and machine learning and artificial intelligence (AI). Training will be tailored to the student’s academic background and existing skills, allowing them to strengthen their Earth science knowledge or computational expertise as needed. Key areas of training will include:
– Domain knowledge: Develop an understanding of Earth surface processes, geomorphology, geospatial data science, and the principles of machine learning and AI.
– Programming and geospatial analysis: Gain proficiency in Python, geospatial data processing, and the analysis of large-scale Earth observation datasets.
– Machine learning: Develop practical skills in machine-learning algorithms, model evaluation, spatial cross-validation, and model interpretation.
– Geospatial foundation models and AI: Learn to work with geospatial foundation models and investigate their potential to improve our understanding of Earth surface processes.
– Scientific research and communication: Strengthen skills in scientific writing, data visualisation, conference presentations, and open and reproducible research practices.

Training activities activities will include:
– Participation in postgraduate research training workshops organised by the British Society for Geomorphology (BSG).
– Attendance at seminars and workshops in relevant research Clusters within the School of Geographical and Earth Sciences.
– Regular interaction with a dynamics group of early-career and senior researchers working across diverse areas of Earth science and geospatial data science.