A computational framework for labeling spatiotemporal remote sensing datasets

Remote sensing instruments and sensors have made significant progress over the last several decade in spatial, spectral and temporal resolutions. These improvements have led to the collection of synoptic scale data and enabled a variety of new applications. For example, improvements in temporal resolution allows monitoring biomass on a daily basis. Improvements in spatial resolution allows fine-grained classification of urban settlements, damage assessments, and critical infrastructure monitoring. Remote sensing applications have the following characteristics: • Spatio-temporal Grid: The underlying data sets are gridded—with the grid dimensions ranging across 2D (images, 2D flow fields), 3D (volumetric data) and 4D (spatiotemporal data). Unlike standard machine learning approaches where the data is usually only available in the form of high dimensional feature vectors, the presence of a grid affords us the potential to develop techniques that can interpolate each of the features (using differentiable splines) and generate new feature vectors with different (and more desirable) nearest neighbor properties. • Large Volume and Velocity: The underlying volume (terabytes to petabytes) and velocity (gigabytes to terabytes per day) of these applications is very large and are responsible for carrying us into the bigdata regime. Effective processing requires low complexity and multiscale algorithms that can exploit modern parallel architectures with deep memory hierarchies.