Quantizing Space and Time
Fusing time series and images for Earth observation
Quantizing Space and Time: Fusing Time Series and Images for Earth Observation
This work introduces a task-agnostic framework for multimodal fusion of time series and images, enabling cross-modal generation and robust downstream performance for Earth observation applications.
Key Innovation
The framework proposes a unified representation space for spatiotemporal data through:
- Discrete Quantization: Transforms continuous time series into discrete tokens
- Masked Correlation Learning: Aligns discrete image and time series tokens
- Task-Agnostic Design: Enables both cross-modal generation and robust downstream performance
- Scalable Architecture: Handles variable-length time series and different spatial resolutions