Paper Info
Title
Towards Complex-Valued Neural Algorithms for Forest Parameters Estimation from Polinsar Data
Abstract
1. ABSTRACT Although the global amount of carbon stored in the worldwide vegetation is a key parameter for quantitatively characterizing the carbon cycle, cost-effective methods for reliably monitoring the existing vegetal biomass and its variations are still lacking. Backscattering amplitude measurements demonstrated their potential in yielding above ground biomass density up to a satura- tion limit, which depends on polarization and, especially, on wavelength. C-band saturates rapidly, while L-band, responding primarily to larger branches, has somewhat extended the saturation limit. However, estimating the biomass of denser forest stands still appears to be beyond the reach of the present amplitude-based techniques. Also the phase of the backscattered field carries information on the biomass, given the scattering and extinction occurring locally within the soil-canopy system, hence interferometric images, in principle, can be converted into biomass maps. The complexity of the wave-canopy interac- tions, dependent also on polarization, makes the retrieval of forest parameters from interferometric (PolInSAR) observables not straightforward. PolInSAR retrieval algorithms are commonly based on an optimal estimation (OE) approach, which computes the sought forest parameters by minimizing the difference between measurements and simulations. Thus the procedure requires an accu- rate coherent scattering model that relates the PolInSAR measurable, i.e., the scattering matrices or derived quantities (1), to an intermediate parameter, i.e., the average tree height. Then the biomass density can be estimated from tree height using the canopy-specific allometric equations. This retrieval scheme is hampered by some drawbacks. Among them, beyond the attain- able accuracy, the convergence of the OE algorithm is not guaranteed, depending on its initialization and/or on the choice of the cost function to be minimized. Another disadvantage derives from the OE iterative scheme, requiring the coherent scattering model to simulate the observables for a number of values of the forest parameters, what makes the procedure extremely time consuming. Neural network (NN) algorithms have proven to be an effective tool for inverting measurements in remote sensing (2, 3, 4), and can perform the estimations in real time. NN can be viewed as a mathematical model composed of nonlinear parallel computational elements processing real-valued quantities, i.e., amplitude and phase separately. Complex Valued Neural Networks (CVNN) have been proposed recently (5), able to handle complex values, that appear suitable to cope with complex quantities as the PolInSAR observables essentially are. Such algorithms, in combination with an electromagnetic model, 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 in the iterative estimation procedure or in extensive multiple computation, once properly trained. This contribution concerns the design of a CVNN algorithm for retrieving tree height and extinction coefficient, with the final purpose of estimating above-ground biomass from PolInSAR data. 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 (6), at L-band and at an incidence angle of 21 degrees. The 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. The computation time for generating the required extensive set of simulations has been dras- tically reduced by utilizing a Direct Neural Network (DNN), both real- and complex-valued, as a direct scattering algorithm. In fact, once trained by a suitably selected set of input-output vector pairs generated by the coherent physical scattering model, the DNN can produce new pairs in near real time. As far as the inversion problem is concerned, given the essentially nonlinear
Year
DOI
Venue
2008
10.1109/IGARSS.2008.4779074
Geoscience and Remote Sensing Symposium, 2008. IGARSS 2008. IEEE International
Keywords
Field
DocType
environmental factors,forestry,geophysical signal processing,geophysical techniques,neural nets,parameter estimation,radar interferometry,radar polarimetry,radar signal processing,remote sensing by radar,synthetic aperture radar,vegetation mapping,CVNN algorithm,PolInSAR data,complex valued neural algorithms,complex valued neural network,forest biomass retrieval,forest parameter estimation,neural net features,neural net training,polarimetric interferometric SAR data,radar backscattering data,Forest biomass,complex valued neural networks,polarimetric interferometric synthetic aperture radar (PolInSAR)
Radar,Algorithm design,Polarimetry,Computer science,Synthetic aperture radar,Remote sensing,Backscatter,Estimation theory,Artificial neural network,Neural algorithms
Conference
Volume
ISBN
Citations 
2
978-1-4244-2808-3
0
PageRank 
References 
Authors
0.34
2
4
Name
Order
Citations
PageRank
Emanuele Angiuli172.82
Fabio Del Frate250872.43
Barbara Polsinelli300.34
Domenico Solimini46515.10