A Study of Tropical thin Cirrus Clouds with Supervised Learning

ABSTRACT Accurate knowledge of the temporal frequency and spatial extent of optically thin cirrus is crucial to climate feedback analysis. Current global warming theory asserts that when the atmospheric concentration of CO2 increases, the outgoing longwave radiation at non-window wavelengths is reduced. If the Earth's net radiative balance is to remain stable, ground temperatures must rise in response, thereby increasing thermal emission to space. Current models do not account for subsequent changes in cloud cover, because this aspect of the climate feedback system is so poorly understood. One possible response of the cloud-climate feedback process is an increase in the global occurrence of thin cirrus clouds, driven by the increase in longwave cooling in the upper troposphere that results from higher CO2 concentrations. Exacerbating the difficulty of assessing the situation is the fact that passive remote sensing instruments cannot reliably detect cirrus clouds with optical depths less than ~0.3, because these clouds do not reflect enough sunlight to create a sufficient contrast with the Earth's surface. Now, however, the presence of thin cirrus can for the first time be accurately detected and systematically monitored by the combination of active and passive sensors onboard the CALIPSO satellite. Nevertheless, the data record is still quite limited, as CALIPSO has been in orbit for only 16 months. We have therefore initiated a multi-platform data fusion study to establish a methodology for extending the limited set of CALIPSO measurements to the existing 30-year record of passive remote sensing data, and thus improve our understanding of cloud feedback mechanisms. Using nighttime data from the first 10 days in April 2007 as a training set, we applied a general regression neural network (GRNN) to collocated samples of sea surface temperature (SST) reported by AMSR, brightness temperatures (BT) from the CALIPSO imaging infrared radiometer (IIR), and optical depths (OD) derived from the CALIPSO lidar measurements. The result is an accurate mapping of the optical depths derived from the active sensors to the brightness temperatures computed from the passive sensor measurements. Applying the trained network to this combination of passive sensor parameters, optical depths as small as 0.1 can be reliably retrieved. The relative uncertainties in the retrieval are reasonable, and can be improved significantly by use of a much larger training set.