Premium
Neural networks for the retrieval of water vapor and liquid water from radiometric data
Author(s) -
Frate Fabio Del,
Schiavon Giovanni
Publication year - 1998
Publication title -
radio science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.371
H-Index - 84
eISSN - 1944-799X
pISSN - 0048-6604
DOI - 10.1029/98rs02133
Subject(s) - artificial neural network , computer science , ceilometer , normalization (sociology) , pruning , linear regression , algorithm , inversion (geology) , range (aeronautics) , data mining , remote sensing , artificial intelligence , machine learning , materials science , geology , cloud computing , paleontology , structural basin , sociology , anthropology , agronomy , composite material , biology , operating system
This paper investigates the potentiality of neural networks for the retrieval of integrated water vapor and integrated liquid water from data simulating the measurements of different groundbased microwave radiometers, with the optional addition of a laser ceilometer. The reliability of the neural network algorithms was evaluated comparing their performance with that achievable by linear regression techniques with the same data sets. The obtained results showed that the neural inversion provides a more accurate estimation of the parameters to be retrieved, especially in the case of strong nonlinearities. In addition, neural networks succeed in exploiting the information given by the ceilometer significantly better than linear regression. The advantages shown by the neural inversion led the investigation further, aiming at the optimization of the architecture of the net and focusing on the number of processing units and connections. An optimum range for the choice of the number of neurons to be inserted in the net has been determined, and ineffective connections have been removed via pruning algorithms. A fault tolerance analysis was also performed, which pointed out other interesting properties of the neural retrieval procedure. In conclusion, the results of this simulation indicate that neural networks, if compared with linear regressions, are more suitable for cases with stronger nonlinearities and are more flexible and robust algorithms. Moreover, their performance can improve when the number of units and connections is optimized.