A Compressive Sensing and Deep Learning Approach for Data Fusion of Multi-Satellite Observations

Abstract

Atmospheric carbon dioxide concentration in the last few years has been increasing abnormally. This increase appears to be responsible for a number of changes in the Earth’s climate. Monitoring the carbon cycle is highly important as carbon plays an important role in climate change.Thus, a close analysis of the concentration of carbon dioxide on Earth is needed to help in understanding the potential feedback of the carbon cycle on vegetation, floods, droughts, hurricanes, and the number of species disappearing. In this study, we analyze multi-year observations from two orbiting satellites measuring total column CO2 concentration (XCO2). In particular, we wish to study methods for combining the two measurement systems by fusing the data from the two satellites, OCO-2 and GOSAT. We will estimate the daily changes in CO2 concentrations for a four-year period over 3 geographic regions varying in latitude. The 3 sites are (i) continental United States and (ii) the equatorial Amazon basin consisting mostly of Brazil and Bolivia, (iii) a polar region extending from Alaska an upper portion of western Canada at the same latitudes. These 3 regions at different latitudes represent the different latitudinal variances of CO2. The breadth of longitude yields greater satellite data land surface coverage. We show in this thesis that one can apply the theory of Compressive Sensing (CS) for fusing multiple data sets while maintaining the temporal frequencies of both. We apply the CS method of data fusion to the combination of OCO-2 and GOSAT satellite measurements of CO2 over 4 years to study the variability of the carbon cycle at low, mid and high latitudes over the Americas. The CS uses a small sample of data from the merging of both data sets to reconstruct a new data set with greater accuracy than each of the separate data sets. A requirement for CS is that the input data be sparse, which we enable through a wavelet transform of the merged dataset. Furthermore, we use a Convolution Neural Network (CNN) model to remove the noise in the reconstructed output. We show that data fusion using CS and CNN can produce a reconstructed data set with less noise and data gaps. This result provides a data set for the analysis of CO2 concentration, that reduces the noise in the residual reconstructed data set. Our goal is to show that the reconstructed data set obtained from two satellites is more accurately compared with US Ameriflux stations than each of the separate satellite measurement systems.