Application of neural networks to atmospheric pollutants remote sensing

Abstract

Infrared remote sensing is an extended technique to measure ”in situ” atmospheric pollutant gas concentration. However, retrieval of concentrations from the absorbance spectra provided by technique is not a straightforward problem. In this work the use of artificial neural networks to analyze infrared absorbance spectra is proposed. A summary of classical retrieval codes is presented, highlighting advantages and important drawbacks that arise when these methods are applied to spectral analysis. As an alternative, a neural network retrieval approach is suggested, based on a multi layer perceptron. This approach has been focused to the retrieval of carbon monoxide concentration, because of the great environmental importance of this gas. Absorption overlapping of atmospheric gases such as carbon dioxide, nitrous oxide or water vapour is one the most important problem in the retrieval process. The training dataset has been generated with special care to overcome this aspect and guarantee a successful training phase. Results obtained from the ANN method are very promising. However, high retrieval errors have been found when ANN method is applied to experimental spectra. This fact reveals the need of a deep study of the instrumental parameters to be included in the model.