Applied machine learning tools for earth systems

Thesis

Every day, hundreds of petabytes of Earth Observation data is created, capturing the information to describe the surface and atmosphere conditions of our planet. To effectively ingest, synthesize, and make sense of the enormous amount of data, Artificial Intelligence (AI) poises itself as a unique tool to extract previously unknown relationships in the data across spatial and temporal domains for a diverse breadth of applied downstream tasks. We define this area of research as application-driven AI to support scientific understanding. Continuously, machine learning algorithms designed in entirely methods-focused research fall short when used directly for applications. In these contexts, it is not enough to only improve predictive performance of a target variable in a scientific domain, we need to further extract meaningful insight via AI model predictions and interpretations, provide confidence metrics to support our findings, and create actionable information for end users. Our work contributes to this through three interrelated pillars a part of our research agenda, falling underneath the common theme of developing applied machine learning tools for Earth’s Systems. We target our research goal through the lens of climate modeling, disaster preparedness, and humanitarian strength building in developing regions.

In our climate modelling chapter, we build an uncertainty-aware deep learning emulator to capture physical model bias of surface ozone. We leverage uncertainty quantification methods to compare and contrast spatio-temporal regions of high uncertainty per methodology, to assist with the down-selection of ground-based ozone monitoring stations for bias correction. We demonstrate improved bias modelling capabilities through leveraging our pythonic geospatial data processing package, which we also utilize in our further chapters.

In our disaster preparedness chapter, we use open-source EO data to create an end-to-end sub-seasonal landslide impact forecasting tool for Nepal. We demonstrate operational capabilities by deploying the tool with the Nepal Humanitarian Country Team to support disaster preparedness, and provide a spatial and temporal analysis of landslide impact forecasting capabilities using the tool.

In our humanitarian strength building chapter, we present our work on leveraging geospatial foundation models for predicting internet connectivity in the Global South, to support the United Nations Children’s Fund initiative to connect every school to the internet by 2030.

Our body work presents unique AI tools to support Earth’s Systems applications utilizing data-driven approaches with the fundamental motivation to improve life on Earth.

Authors: Doerksen, K

• DOI: 10.5287/ora-pryam1k9y
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning; earth observation
• Subjects: Machine learning