A Study On The Retrieval Of Clear-sky Atmospheric Humidity Profiles Using Artificial Neural Network Algorithm

This thesis is a attempt to retrieve the atmospheric humidity profiles from hyperspectral Atmospheric Infrared Sounder (AIRS) data using artificial neural networks(ANN).Get the conclusion that ANN is slightly better than Eigenvector Regression Algorithm(ERA) on this problem and show strong ability on dealing with the nonlinear problem.The key and aporia is how to build a neural network with supreme generalization ability to achieve high precision.BP network,which is the most widely used,is Selected.Proposes a solution according to over fitting and strong owe set.After test, confirmed the numbers of hidden nodes,using conjugate gradient learning algorithm and tangent transfer function, established network structure.The data sets use the group of the global clear sky training profiles sample collection from CIMSS (Cooperative Institute for Meteorological Satellite Studies, University of Wisconsin-Madison) and the simulated AIRS brightness temperature which derived from SARTA (Stand-Alone Radiative . Transfer Algorithm) forward pattern.First,inspection accuracy of atmospheric humidity profiles between ANN and ERA using inspection sample.It is find that the regression results of ANN are better than ERA at all pressure layers.Second,use AIRS live observation data instead of simulation data.For example, pick up the data which covered Chinese region on the sep.6.2002,compared the two regression results to the corresponding AIRS Level 2 Atmospheric products and ECMWF objective analysis field.No obvious advantages in both algorithm form general RMSE(root mean square error) curve of atmospheric humidity profiles when determine regression coefficient using the global scope samples.But ANN is obvious better than ERA,and its RMSE is less than ERA at all flight level using Chinese region samples.In the end of the paper analyzed the reasons of why ANN Can't obviously improve precision greatly.Puts forward the direction of further research.