Exploring Spatially Distributed Deep Learning Models for Global Gravitational Mapping

A broad range of geoscience applications utilize gravity measurements maintained by the Office of Geomatics at the National Geospatial-Intelligence Agency (NGA), including Earths Gravitational Model. NGAs measured gravity database collection is the largest in the world, yet there are gaps where gravity data is not known due to difficult to reach terrain or geopolitical constraints. Currently, a series of forwarding modeling techniques are used for estimating gravity in areas with little to no collected gravity points. This results in inconsistent mapping products, both in terms of resolution and quality. We are approaching these challenges from a data-driven perspective and aiming for global coverage gravitational maps. We propose to leverage advances from machine learning to model highly nonlinear and complex relationships between the critical input signals and the example gravity anomaly measurements. However, we argue that generating a single machine learning model with highly heterogeneous features might not be suitable to capture the complex variabilities determining the gravity signature, as variability in data quality and feature heterogeneity might greatly impact model generalization. %Joined with the Office of Geomatics at the National Geospatial-Intelligence Agency, We seek to leverage ReSFlow, a stratification dataflow, to partition global scale data inputs into rather spatially distributed homogeneous subsets and aim for optimizing ML models trained with partitioned subsets. Those tailored models are especially targeted for specific forms of variation and are developed and deployable wherever these conditions or similar context are found across different geographies. With preliminary results, we have demonstrated that the ML models trained with partitioned data have shown promising results as compared to the monolithic ML model. We also investigate the importance of input features that we used to predict the gravity model performance by visualizing the model learning progress and feature relevancy and providing the insights for bucket specific model outputs and their importance.