Retrieval Of Ocean Winds Form Ers-1/2 Scatterometer And Sar Data Using Natural Network.
K.S.Chem. J.T.Wang * and A.J.Chem.
Center for space and Remote sensing research
*Department of Atmospheric University
National Central University
Chung-Li, Taiwan
E-mail:dkschen@csrsr.ncu.edu.tw
Abstract
This paper presents the retrieval of the ocean winds observed from C-band microwaver scatterometer onabord ERS-1 satellite platform I April 1994. the study area was located at west Pacific water area (~ 17n , 124E~25N, 135E). a neural network is adopted to implement the inversion of a geophysical model function which relaters the scatter meter measurements of normalized radar cross section to surface wind speed and direction. To illustrate the functionality of the neural network, a set to wind fields were generated by means of Monte Carlo simulations. At each sample point, the 3ind speed and direction are obtained. Then, a geophysical model function proposed by ESA ( European Space Agency ) was used to produce the simulated normalized radar cross section at three pointing antennae of scatterometer according to the ERS -1 configuration. As a result neural network was constructed as having four imputes nodes accepting three NRSCs and an incident angle at nid-beam antenna, and two output nodes representing the inverted ( and desired ) wind speed and direction. Network training was accomplished by the input output pairs which are randomly selected from the database of simulated wind fields. The effectiveness of the neural network as an inverse transfer function was validated. Applying this well trained neural network to ERS-1 data was presented. Comparisons with traditional optimization method concludes that satisfactory results were able to obtain by the proposed neural network approach. Wh4en makes use of spatial information, improvement of the retrieval accuracy can be obtained. The use of SAR to derive ocean winds is driven by the demands of fine spatial resolution such as in the coast regions. A pair of SAR image data were acquired within the study area of scatterometer with very closely in time. Atmospheric boundary layer rolls ere observable on the image, thus the wind direction was able to determine. The CMOD-4 model was then applied to estimate the wind speed, following a radiometric calibration. No buoy data was available to quantitatively conclude our results. Nevertheless, reasonable agreements were obtained when comported to scatterometer.
Introduction
Scatterometeris microwave radar with capability of relating measured normalized radar cross section to wind speed and direction over the ocean surface. The principles of the scatterometer measuring the wind lies in the fact that the microwave radar echo from ocean surface are dependent on amplitude and density of the waves. These waves are related to wind. In mocrowave region, the Bragg wavelength is in the water wave capillary-gravity waves which are generated by wind [ Moore and Fung, 1979]. The strength of the radar echo also determined by the radar frequency, po9larization and incident angle, as described by radar equation. Hence, scatterometer provides an indirect means of wind field measurements. How-ever , the interaction of the radar signal and owned-roughed sea surface are complicated. For example, the mechanism of the generation of waves by the wind is poorly understood. The relation between the wind at 10 m above the surface and momentum flux into the ocean is dependent on the surface wave structure, the layer strification , and the magnitude of the wind. In addition, recent experiential observations reported that the NRSC may change significantly across a sea surface temperature front . The success and value of scatterometer, therefore, requires a good relationship between the scatterometer signal na surface wind. Practically speaking, there are two basic problems that need to solved: a transfer function that relates the NRCS to wind field, and a method that inverts wind field form NRCS. Mathematically,
s0 = G(V,q, f,p) (1)
V= G(s0 , q, f,p) (2)
Where
s0 is NRCS measured by scatterometer, V is surface wind vector ( speed and direction ); ? is incident angle, f is radar frequency, and P is polarization. Note the radar system parameters are know but may subject to nose contanimation.
In the above, G is known as geophysical model function and g is inverse transfer function which is usually a multivalued function. Hence, several measurements of
s0 form different azimuth angle must be used to estimate the wind vector . conventional approach to wind estimate involves formulating a cost function from the scatterometer measurements and then minizing it to obtain estimates of wind field. Local minima are usually appear due to the nature of geophysical model function . this leads to the some called aliasing or ambiguity. Since dealiasing or ambiguity removal relies on the information not contained in
s0 measurements, it is not pursed here. In this paper, we are primarily concern with the inverse transfer function g. Because it is nonlinear5 and the exact character of the nonlinear behavior may not be known a priori. Neural network offers a highly flexible, yet accurate, alternative to implement the transfer function. In the next section, the scatterometer models are first mentioned to facilitate the sections that follow. Section 3 presents the neural network approach to implementing the inverse transfer function, in particular, a dynamic learning neural network (DLNN) trained by a Lalman filtering technique is applied. The input-output structure of the network and learning algorithm are given. The validation and verification the network are accomplished through numerical simulation.
