Deep Learning to Address Data Sparsity in Climate Change Monitoring

Abstract

Data sparsity is a challenge in climate change. There is a scarcity of flux towers that measure climate quantities such as GPP and Reco. The motivation is to predict values for GPP and Reco with no labeled data since GPP and Reco are important quantities for carbon cycle predictions. The goal is to leverage knowledge from well-observed entities to improve predictions for unobserved ones, and to incorporate domain knowledge into machine learning models for better generalization. We present a deep-learning approach called Concatenation Long Short-Term Memory (CT-LSTM) to enable predictions for GPP and Reco using minimal labeled data. CT-LSTM leverages drivers (meta features/ characteristics) to transfer information from well-observed sites to unobserved ones which can then later be applied to these unobserved sites.