|Paper title||Uncertainty estimation via Neural Networks : A Quantile Regression Neural Network for the retrieval of greenhouse effect gases and its associated uncertainty|
|Form of presentation||Poster|
In the scope of remote sensing retrieval techniques, methods based on deep learning algorithms have gained an important place in the scientific community. Multi-Layer Perceptron (MLP) Neural Networks (NN) have proven to provide good estimates of atmospheric parameters and to be more performant than classical retrieval methods – e.g. Optima Estimation Method (OEM) – in terms of computational cost and processing of non-lineal models.
However, the most important drawback of current classical MLP techniques is that they do not provide uncertainty information on the retrieved parameters. In the atmospheric retrieval challenge, not only the quantitative value of the computed parameter is important, but also the incertitude associated with this estimation. The latter is essential for the exploitation of scientific products, for example, its utilisation in analyse/forecasting systems of the atmospheric composition or dynamics. In order to come up with a solution to the incertitude estimation issue, new MLP NNs have been recently developed – e.g. Bayesian Neural Networks (BNN) and Quantile Regression Neural Networks (QRNN) –.
The French National Centre of Spatial Studies (CNES) is therefore interested in developing and proving the feasibility of NN methods for the modelling of the incertitude associated with atmospheric variables, and more specifically, in the retrieval of greenhouse gases – e.g. CO2 content – obtained from infrared hyperspectral sounding instruments such as IASI, IASI-NG or OCO-2.
To this end, a QRNN (Quantile Regression Neural Network) has been implemented in order to estimate the mid-tropospheric CO2 distribution probabilities for a synthetic set of brightness temperatures corresponding to selected channels of IASI and AMSU. These sets are representative of a wide range of atmospheric situations in the tropical zones of the globe, including extreme events.
The present QRNN is then able to retrieve the predicted probability intervals of the tropical mid-tropospheric CO2 column – in this case, 11 quantiles positions ranging from 0.05 to 0.95 –. Validations show a robust and well-calibrated neural network with an accurate retrieval of the CO2 content and coherent associated incertitude estimation for a wide set of brightness temperatures corresponding to a CO2 range between 396 and 404 ppmv. Indeed, the implemented QRNN is able to associate a greater uncertainty to the most biased CO2 estimations. This performance criteria is of great importance for later applications that take advantage of retrieval/inversion products, allowing for the filtering of the doubtful – i.e. uncertain – estimates and thus the obtaining of more accurate results – e.g. better assimilation products –.