AUTOMATIC FEATURE SELECTION FOR COMBINED IASI / GOME-2 OZONE PROFILE RETRIEVAL

The Neural Network Ozone Retrieval System (NNORSY) version 2 uses hierarchical neural networks to retrieve the atmospheric ozone profile from the combined spectra of the GOME-2 and IASI instruments on METOP. Formerly, we used a similar method with a single neural net to retrieve ozone from several UV/VIS sensors. The reason for originally developing a hierarchical network model lies in the overwhelming complexity of a joint retrieval paired with very high computational costs, which made it impossible for a simple 3-layer feed-forward neural net to learn the highly non-linear mapping from combined UV/VIS and IR spectra to ozone profiles. Recent advances in computing technology have now made the hierarchical approach mostly obsolete, as our new network training software accelerates learning by a factor of more than 100 using graphics processing units (GPUs). This enables us to use Deep Neural Networks (DNNs) with five or more layers, which can learn the higher order correlations between input spectra and target profiles without a semi-automated hierarchical network construction step. Due to these improvements in automation and speed, it was possible to implement a meta-learning approach using a state-of-the-art exploration technique named PGPE (Policy Gradients with Parameter-based Exploration; Sehnke et al. 2010), a recently developed Reinforcement Learning algorithm which has been found to have numerous advantages over evolutionary strategies. This algorithm automatically trains many networks with differing input features and architecture until the best combination is found. This leads to an improvement of retrieval accuracy especially in the troposphere and UTLS region. In this paper we present the new approach and compare the former more physically based channel selection for GOME-2 and IASI with the results of the automated feature selection process. While many spectral channels come out as expected, there are also several interesting differences which we are going to discuss.