Software has been developed that mitigates anomalously low CO2 retrieval values, due to low-lying clouds, for improved CO2 sensing accuracy. Using machine learning techniques, hand-labeled data was used to train intelligently a multivariate least-squares filter using only five input-level-one features that detect low-lying clouds over ice with more than 99% accuracy. These input features are simple arithmetic operations on the input spectrum such as max, min, std, and mean.
Performing a full CO2 retrieval from the GOSAT (Greenhouse gases Observing SATelite) data soundings is time-intensive (at 20 minutes per sounding). As ACOS (Atmospheric CO2 Observations from Space) is a preparatory task for the Orbiting Carbon Observatory-2 (OCO-2) coming mission in which data volumes will be much higher, using the GOSAT environment to test and develop intelligent filters that can predict failed or useless retrievals before significant computational power has been wasted is a high priority.