Printable Version of this PageHome PageRecent ChangesSearchSign In
Abstract: Validation and Interpretation of Satellite CO2 Data using Neural Networks. By Dan Knights, 12/10/2007.

Increases in the level of anthropogenic CO2 emissions over the last century are having significant effects on the planet's climate. The international community generally agrees that there are clear indications that the overall increase in CO2 in the atmosphere is having a markedly deleterious effect on the global climate. The global interaction of thousands of CO2 sources and sinks is highly complex, however, and we currently have only several hundred sites worldwide for monitoring surface and tropospheric CO2 flux [1]. The cost of ground-based CO2 level observation is prohibitively high, but recent advances have been made in measuring atmospheric CO2 levels via satellite-mounted sensors, such as the Atmospheric Infrared Sounder (AIRS). Recent studies have shown that the measurements taken by these new sensors can be accurate at certain times, and inaccurate at others. They are also highly sensitive to systematic measurement errors.

This thesis proposes a novel method both for validating the satellite-based sensors, and for interpreting the data that they produce. The method will employ artificial neural networks (ANN's) to provide a meaningful interpretation of the raw satellite data. Portions of the satellite data that are known to contain measurement error, time lag, or discontinuities as well as portions that are known to be correct will be coupled with certain known ground-based CO2 measurements to provide training data for the ANN. The ANN will then be used to make predictions of ground-based measurements based on the raw satellite data. The ANN will be tested using a different set of satellite data containing both good and bad measurements.

The thesis will provide an analysis of the performance of the ANN as a predictive model for conditions of time lag, discontinuities, and systematic measurement error. The predictions made by the ANN will be compared with other models, such as regression, Dynamic Time Warping (DTW), and regression coupled with DTW. Due to the common success of ANN's in the field of climate-related predictive modeling [2], ANN's are expected to outperform other methods of interpretation and validation of the satellite-based measurements.

[1] Y. Tiwari, “Constraints of Satellite Derived CO2 on Carbon Sources and Sinks,” Ph.D. Thesis, Max-Planck-Institute for Biogeochemistry, 2006.

[2] G. Chattopadhyay-Bandyopadhyay, "Artificial Neural Network versus autoregressive approach: Prediction of total ozone time series," Model Assisted Statistics and Applications Volume 2, Number 3, p. 107-120. 2007.

Last modified 10 December 2007 at 4:00 pm by danknights