Complex-valued Neural Network Algorithms for Forest Parameters Retrieval and Classification from Polarimetric SAR Data
Emanuele Angiuli(1), Fabio Del Frate(1) and Domenico Solimini(1)
(1) University of Rome Tor Vergata, Via del Politecnico, 1, 00133, Rome, Italy
Complex Valued Neural Networks (CVNN) have been proposed recently , able to handle complex values, that appear suitable to cope with complex quantities as the PolSAR observables essentially are. Such algorithms learn the computational relationships directly from the complex-valued inputs during the training phase and, once trained, are able to process large sets of experimental data in very short time.
Moreover, the interpolation properties of NN are known. Hence, they can replace the time-consuming physical scattering model or traditional classification techniques, once properly trained.
This contribution concerns the design of a CVNN algorithm for forest parameters retrieval and classification from PolInSAR data.
For the retrieval maps, the design procedure has been based on sets of simulations of scattering matrices carried out by the coherent scattering model implemented in the POLSARPRO tool , at L-band.
Simulations include the effects of random variations of soil moisture and of surface roughness and consider joint statistical distributions of tree height and extinction coefficient. As far as the inversion problem is concerned, given the essentially nonlinear nature of the NN computations, in principle the algorithm can handle the raw scattering matrix without any preliminary decomposition . To elaborate on this feature, we checked the results of our inversion procedure with and without scattering matrix decompositions for a selected sub-set of simulations. The values retrieved by our algorithm are compared with those yielded by the optimal estimation approach to analyse and discuss robustness of the algorithms and retrieval accuracy, also with reference to the range of variation of the forest parameters.
Instead, for classification maps we have considered real data from ALOS-PALSAR sensor. In this case we trained the net considering different areas of the data and associating to them different output
class for the network. After an extensive training, the best network topology was used to classify all the data collected. Classification results are then discussed and analyzed through the traditional accuracy
 A. Hirose, Complex-Valued Neural Networks: Theories and Applications, Springer-Verlag, Berlin,2006.
 “Polarimetric SAR Data Processing and Educational Tool (POLSARPRO) v3.1 site,” [Online] Available at: http://earth.esa.int/polsarpro/, Jan. 2008.
 S. R. Cloude and E. Pottier, “A review of target decomposition theorems in radar polarimetry,” IEEE Trans. Geosci. and Remote Sens., vol. 34, no. 2, pp. 498–518, 1996.