ORBIT (Ordering Based Information Transfer): A Physics Guided Machine Learning Framework to Monitor the Dynamics of Water Bodies at a Global Scale

Abstract

Predictive learning approaches along with vast amounts Earth Observation data offer a great opportunity to track changes on the earth's surface. However, due to data quality issues (sensor anomalies and atmospheric disturbances) and heterogeneity in the land surface, even state of the art machine learning algorithms perform poorly when applied on a global scale. Furthermore, due to inherent trade-off in sensor design, a single source does not provide both high spatial and temporal resolution required by various scientific applications. This thesis focuses on developing new machine learning algorithms that can leverage physical principles governing geo-physical processes to overcome these challenges, in the context of monitoring surface water changes at global scale. The thesis introduces a new framework, ORBIT (Ordering Based Information Transfer) that uses an implicit ordering constraint among instances to address the aforementioned challenges. For this application, the topography (the elevation structure) enforces such an ordering. This elevation constraint, however, is not available explicitly in almost all the cases. This thesis introduces a new rank aggregation approach to infer the inherent ordering from the noisy labels. This thesis also introduces a new approach that makes use of this elevation constraint to enforce temporal consistency in surface area variations of water bodies. Finally, this thesis introduces a new approach to downscale low resolution land/water masks to a higher spatial resolution using elevation ordering available at high resolution.