A highly dynamic learning (DL) neural network is developed and applied to perform the inversion of rough surface parameters: dielectric constant, surface rms height, and correlation length. The network training scheme is based on the Kalman filter technique which lends itself to a highly dynamic and adaptive merit during the learning stage. The training data sets utilized were obtained from the Integral Equation Model (IEM) which has a wide range of frequency. The training speed of the network is found to be much faster than the back-propagation (BP) trained multi-layer preceptron (MLP) with the same degree of accuracy. When applied to invert the surface parameters, the DL network shows a very satisfactory result in terms of learning time and process accuracy which thus enhances its potential applications to remote sensing of rough surfaces.
