Combined Ozone Retrieval from METOP Sensors using Meta-Training of Deep Neural Networks
Felder, Martin; Sehnke, Frank; Kaifel, Anton
Zentrum für Sonnenenergie- und Wasserstoff-Forschung (ZSW), GERMANY

The newest installment of our well-proven Neural Network Ozone Retrieval System (NNORSY) combines the METOP sensors GOME-2 and IASI with cloud information from AVHRR. Through the use of advanced meta-learning techniques like automatic feature selection and automatic architecture search applied to a set of deep neural networks, having at least two or three hidden layers, we have been able to avoid many technical issues normally encountered during the construction of such a joint retrieval system. This has been made possible by harnessing the processing power of modern consumer graphics cards with high performance graphic processors (GPU), which decreases training times by about two orders of magnitude.

The system was trained on data from 2009 and 2010, including target ozone profiles from ozone sondes, ACE-FTS and MLS-AURA. To make maximum use of tropospheric information in the spectra, the data were partitioned into several sets of different cloud fraction ranges with the GOME-2 FOV, on which specialized retrieval networks are being trained. For the final ozone retrieval processing the different specialized networks are combined.

The resulting retrieval system is very stable and does not show any systematic dependence on solar zenith angle, scan angle or sensor degradation. We present several sensitivity studies with regard to cloud fraction and target sensor type, as well as the performance in several latitude bands and with respect to independent validation stations. A visual cross-comparison against high-resolution ozone profiles from the KNMI EUMETSAT Ozone SAF product has also been performed and shows some distinctive features which we will briefly discuss.

Overall, we demonstrate that a complex retrieval system can now be constructed with a minimum of machine learning knowledge, using automated algorithms for many design decisions previously requiring expert knowledge. Provided sufficient training data and computation power of GPUs is available, the method can be applied to almost any kind of retrieval or, more generally, regression problem.