Deep Learning to Address Data Sparsity in Climate Change Monitoring

Abstract

Accurate and continuous monitoring of terrestrial carbon fluxes, such as Gross Primary Production (GPP) and Ecosystem Respiration (Reco), is needed to fully understand and monitor the impacts of climate change. However, the existing observation systems, like the eddy covariance flux towers, are spatially sparse and very unevenly distributed, which leaves significant gaps in monitoring climate variables, especially in tropical regions, which are regions of critical importance for the global carbon cycle. This thesis addresses the challenges of data sparsity in climate change monitoring by using deep learning to predict GPP and Reco in unmonitored locations.

A Concatenation Long Short-Term Memory (CT-LSTM) approach is presented in this work that uses temporally continuous inputs, vegetation indices from remote sensing (e.g., LAI), and static spatial metadata (such as latitude, longitude, and climate zone classification) to predict GPP and Reco in a zero-shot fashion. The model analysis is done using ensemble methods to provide uncertainty estimates.

Experiments using CT-LSTM demonstrate that the approach presented in this work outperforms the baseline model, MetaFlux [1]. The results show the model’s potential for scalable, data-driven climate monitoring using machine learning techniques. This thesis contributes to a step toward more equitable and efficient climate observation frameworks, which are needed in this era of global climate change.