Rethinking the grid: Towards less distorted imagery and AI
Satellite imagery is traditionally stored and processed on rectangular grids. However, the widespread usage of such grids has normalized their inherent distortions, particularly near the poles. Previous attempts to address this issue, such as employing multiple local projections like the UTM-based Sentinel 2 L1C grid, have led to inefficiencies, including a significant increase in data volume (~30%) due to overlaps that need to be stored, downloaded, and processed. Additionally, there is a lack of a unified global indexing system and the choice of pixel cell shape, which further complicate the analysis.
In this keynote talk, we advocate for a paradigm shift towards Discrete Global Grid Systems (DGGS) to mitigate these challenges. DGGS tessellate the Earth's surface with hierarchical cells of equal area, minimizing distortion and reducing loading time of large geospatial datasets. This approach would greatly improve spatial statistics and convolutional Machine Learning models, where accurate representation of global phenomena is paramount at a global scale.