A Bottom-Up Sampling Strategy for Reconstructing Geospatial Data from Ultra Sparse Inputs.

Working with observational data in the context of geophysics can be challenging, since we often have to deal with missing data. This requires imputation techniques in pre-processing to obtain data-mining-ready samples. Here, we present a convolutional neural network (CNN) approach from the domain of deep learning to reconstruct complete data from sparse inputs. CNN architectures are state-of-the-art for image processing. As data, we use two-dimensional fields of sea level pressure (SLP) and sea surface temperature (SST) anomalies. To have consistent data over a sufficiently long time span, we favor to work with output from control simulations of two Earth System Models (ESMs), namely the Flexible Ocean and Climate Infrastructure and the Community Earth System Model. Our networks can restore complete information from incomplete input samples with varying rates of missing data. Moreover, we present a technique to identify the most relevant grid points of our input samples. Choosing the optimal subset of grid points allows us to successfully reconstruct SLP and SST anomaly fields from ultra sparse inputs. As a proof of concept, the insights obtained from ESMs can be transferred to real world observations to improve reconstruction quality. As uncertainty measure, we compare several climate indices derived from reconstructed versus complete fields.