Reconstruct Geospatial Data from Ultra Sparse Inputs to Predict Climate Events

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 approach from the domain of deep learning to reconstruct complete information from sparse inputs. As data, we use various two-dimensional geospatial fields. To have consistent data over a sufficiently long time span, we favor to work with output from control simulations of two Earth System Models, 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 apply a bottom-up sampling strategy to identify the most relevant grid points for each input feature. Choosing the optimal subset of grid points allows us to successfully reconstruct current fields and to predict future fields from ultra sparse inputs. As a proof of concept, we predict El Niño Southern Oscillation and rainfall in the African Sahel region from sea surface temperature and precipitation data, respectively. To quantify uncertainty, we compare corresponding climate indices derived from reconstructed versus complete fields.