On The Selection Of Input Measurements For Neural Network Based Retrieval Algorithms
Michele Federico Iapaolo(1), Fabio Del Frate(1), Fabrizio Rossi(2), Stefano Casadio(3), Sophie Godin-Beeckmann(4) and Monique Petitdidier(5)
(1) Tor Vergata University, Viale del Politecnico, 1, 00133 Rome, Italy
(2) Tor Vergata University, Viale del Politecnico, 1 , 00133 Rome, Italy
(3) Serco Italia, Via G. Galilei, 00044 Rome, Italy
(4) Institut Pierre Simon Laplace, 4, Place Jussieu, 75252 Paris, France
(5) IPSL, 10-12 rue de l'Europe, 78140 Velizy, France
The derivation of atmospheric parameters from satellite remote sensing instruments plays a key-role in monitoring the Earth's atmosphere and understanding the chemical and physical processes therein. However, the atmospheric structure results from a complex interaction between radiative, physical and chemical processes. Neural Networks (NNs) can be a useful tool to face with such complexities. They are composed of many nonlinear computational elements (called neurons) operating in parallel and linked with each other through connections characterized by multiplying factors. This structure makes neural networks inherently suitable for addressing non-linear problems. The derivation of particular rules or statistical a priori information on the data to be processed is not necessary, and the neural networks establish the inverse mapping and the input-output significant relations on the base of data presented to them during the learning phase.
Different NN-based algorithms for the estimation of atmospheric parameters from satellite data have been recently proposed. Since the satellite sensors provide measurements in a very large spectral range, the crucial point in the design of neural algorithms is the selection of the input measurements, which can significantly affect the final retrieval performance.
In this work we focused on the inversion of vertical ozone profiles from GOME radiance measurements. Different methodologies, both automatic and model-based, aiming at the selection of neural network inputs have been analysed, and the obtained results have been critically discussed. According to this analysis, different neural algorithms schemes have been designed for the actual retrieval operation; the final selected topologies have been trained having in input the selected radiance values measured by the sensor, and in output the corresponding ozone profiles provided by the Rutherford Appleton Laboratory (RAL). Once trained, the estimation capabilities of the neural schemes has been extensively validated, either with satellite measurements of the Improved Limb Atmospheric Spectrometer (ILAS) boarded on ADEOS or with lidar observations performed at different stations belonging to the Network for Detection of Stratospheric Changes (NDSC), and the obtained results have been compared.
It has to be reminded that the results of this study can be easily extended to the spectral measurements provided by other satellite instruments, such as SCIAMACHY, boarded on Envisat (2002), or OMI, carried by the EOS-Aura platform (2004).
Keywords: ESA European
Space Agency - Agence spatiale europeenne,
observation de la terre, earth observation,
satellite remote sensing,
teledetection, geophysique, altimetrie, radar,
chimique atmospherique, geophysics, altimetry, radar,