Station Optimization
Multimodal Vision Transformer architectures fusing high-resolution terrain, lidar, and satellite data to optimize where environmental monitoring stations should go.
Where you place a monitoring station determines what you can and can't observe. This project designs multimodal Vision Transformer architectures that fuse high-resolution (3m) terrain, vegetation, airborne lidar, satellite observations, and physically based model outputs to optimize environmental monitoring network design.
Overview
Instead of placing stations by intuition, cost, or accessibility, this project trains models to predict where new monitoring infrastructure will actually reduce forecast uncertainty — treating network design itself as a prediction problem.
Engineering Challenges
Fusing terrain, lidar, multispectral satellite imagery, and model output means reconciling wildly different native resolutions, coordinate systems, and data volumes into a single training pipeline — and doing it across multi-terabyte regional datasets without the preprocessing step becoming the bottleneck.
Architecture
The models train on GPUs using multimodal Vision Transformer architectures that learn joint representations across modalities, targeting high-resolution spatial prediction of where observational value is highest. Training data spans multi-terabyte geospatial datasets, which pushed the data-loading and augmentation pipeline to be as much of an engineering focus as the model architecture itself.
Status
Active research and development, building on the same distributed-training infrastructure used across the rest of this work.
Future Directions
Validating candidate station placements against held-out observation networks, and extending the approach to additional environmental variables beyond snow.