Informing Great Lakes Observing System Design with Convolutional Gaussian Neural Processes
Date:
Using uncertainty-aware machine learning to evaluate environmental observing strategies in the Great Lakes
Overview
Environmental observations are expensive, sparse, and often collected in locations that were chosen for historical or logistical reasons. This raises a fundamental question: if we can only observe part of a system, where should we place new sensors?
In this talk, I presented recent work using Convolutional Gaussian Neural Processes (ConvGNPs) and the DeepSensor framework to inform observing system design in the Great Lakes. These methods learn spatial covariance structures directly from environmental data and provide estimates of predictive uncertainty across space and time.
By identifying locations where uncertainty is highest, uncertainty-aware machine learning can help evaluate candidate observing strategies and highlight regions where additional observations may provide the greatest information gain.
Key Topics
- Great Lakes observing system design
- Convolutional Gaussian Neural Processes (ConvGNPs)
- DeepSensor
- Environmental uncertainty quantification
- Sensor placement and information gain
- Machine learning for environmental monitoring
Take-Home Message
Machine learning can help us explore observing system design choices, but it does not determine what an observing system should optimize for. Effective observing systems require both quantitative tools and human judgment about scientific objectives, stakeholder needs, and practical constraints.