Non-linear inversion of Odin sub-mm observations in the lower stratosphere by neural networks

Book chapter, 2023

Odin is a small satellite employing sub-mm (480-580 GHz) radiometry of atmospheric thermal emission, in a limb sounding mode, to measure global distributions of several species important for ozone chemistry in the middle atmosphere. Although the planned standard data products are based on linear retrieval algorithms, addressing the non-linear effects in some of the Odin observation bands results in an extension of the retrieval altitude range into the lower stratosphere. This paper examines the possibility of applying a technique based on multilayer perceptrons to address the non-linear effects when inverting some of the future Odin data. The study is based on inverting simulated spectra in the 544.2-545.0 GHz band to retrieve O3 profiles by the neural network technique and also by an iterative implementation of the optimal estimation method based on the Marquardt-Levenberg algorithm. First results show that the proposed technique performs in similar terms when compared with the iterative implementation of the optimal estimation method, but offers the advantage of performing the inversions more efficiently from a computational point of view.