Machine Learning with Small-Satellite Imagery for COVID-19 Indicator Analysis

Abstract:

As part of a greater initiative in partnership with NASA to understand the impact of the COVID-19 pandemic on the planet, we developed a machine learning pipeline to detect various economic indicators - airplanes, cars, trucks and ships - within commercial small satellite imagery during the time period of early January to present. Our approach shows how measures enforced to impede the rate of contraction may have potentially correlated with changes in activity amongst the economic indicators through a comprehensive time series analysis, and how those potential changes may be used to infer correlations with varying levels of movement restriction policies implemented internationally. Our models and analysis show how machine learning can couple with other forms of earth observation science to infer rapid insights over time. Specifically in this case, we investigate how detections of various economic indicators might correlate with detected changes in atmospheric and other environmental indicators. Furthermore, the models that we developed were trained using a combination of open source code and benchmark datasets. We will discuss data assimilation methods, model hyper-parameter tuning, scalable inference in the cloud and model retraining feedback loops.