1366 XCO2 Retrieval Using a Neural Network–Based Algorithm from OCO–2 measurements

Wednesday, 15 January 2020
Hall B (Boston Convention and Exhibition Center)
Jaemin Hong, Yonsei Univ., Seoul, Korea, Republic of (South); and J. Kim, W. Kim, Y. Cho, H. Chong, and H. Lim

With the global warming since the Industrial Revolution, it is necessary to monitor the sinks and sources of atmospheric carbon dioxide (CO2), which have been known to have the largest impact on the greenhouse effect. Recently, several satellites with hyperspectral short-wave IR sensor started to operate. Orbiting Carbon Observatory-2 (OCO-2) was launched on September 2014, and has provided about a million of observations every day with resolution and accuracy needed to detect regional scale fluxes. In this study, we developed an algorithm for retrieving column-averaged dry air atmospheric density (XCO2) using OCO-2 measurements based on neural network. On the python Keras-Tensorflow framework, the neural network algorithm consists of 3 layers with 70 neurons for each layer. Cloud mask data of the Moderate Resolution Imaging Spectrometer (MODIS) onboard the Aqua satellite was co-located and used to exclude cloud-contaminated pixels. Measured radiance spectra were normalized by solar spectrum so that normalized spectra have approximately linear relationship with optical depth of the atmosphere. To decrease the number of input variables, the normalized spectra were then multiplied by the small number of the empirical orthogonal functions (EOFs) to yield coefficients for each EOF. The input variables of the algorithm included the coefficients, observing geometry, and time of observation. For validation, the preliminary retrieval result of the algorithm was compared to the Total Carbon Column Observation Network (TCCON) ground-based observations from Nov. 2014 to Dec. 2017 globally. The retrieval results in the northern hemisphere were matched reasonably to the TCCON observations with a very low computational cost. However, the retrieval results in the southern hemisphere showed relatively large error, probably due to small volume of training data compared to the northern hemisphere.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner