RESEARCH COMMUNICATIONS CURRENT SCIENCE, VOL. 91, NO. 10, 25 NOVEMBER 2006 1382 *For correspondence. (e-mail: rmgairola@yahoo.com) Multiparameter microwave retrieval algorithms: performance of neural networks R. M. Gairola*, Samir Pokhrel, A. K. Varma and Vijay K. Agarwal Oceanic Sciences Division, Meteorology and Oceanography Group, Space Applications Centre, Ahmedabad 380 015, India The present study deals with a major aspect of retrieval of oceanic and atmospheric parameters from Tropical Rainfall Measuring Mission-Microwave Imager (TRMM-TMI) channels following the sensitivity studies carried out earlier based on radiative transfer model simulations for rain-free atmospheric conditions over the global tropical oceans. The potentiality of artificial neural network (ANN) for retrieval of geophysical parameters like wind speed, total precipitable water (TPW) and cloud liquid water (CLW) from TMI has been investigated. The radiative transfer simulations of brightness temperatures (TBs) performed for TRMM-TMI frequencies with the inputs from the European Centre for Medium Range Weather Fore- cast (ECMWF) fields of geophysical parameters were used for the constitution of the database of input and output field vectors for the ANN applications. The re- sults show that the neural network algorithm has the capacity to perform retrieval of ocean–atmospheric parameters with good accuracy. Keywords: Algorithm, microwave, neural network. DURING the past several years, the use of satellite-based microwave radiometers for continuous observations of ocean and atmosphere has become common practice. The use of well-calibrated satellite-based microwave radiometers makes it possible to obtain long time series of geophysi- cal parameters. Over the oceans, these parameters include three phases of water: total precipitable water (TPW), cloud liquid water (CLW) and rainfall as well as surface parameters like the sea surface wind speed (WS) and sea surface temperature (SST). The parameters are highly useful in a wide variety of studies of hydrological processes 1 and can improve weather prediction via data assimilation into operational models 2,3 . In order for the satellites to provide the best possible performance in providing most accurate geophysical parameters, it is pertinent to explore the newly emerging techniques in the field of retrievals. The major objective of this paper was to develop a re- trieval algorithm for non-raining oceanic and atmospheric parameters like the WS, TPW, and CLW for the Tropical Rainfall Measuring Mission-Microwave Imager (TRMM- TMI) radiometric channels to a desired accuracy based on artificial neural network (ANN) approach as a prelude to the MADRAS (Microwave Analysis and Detection of Rain and Atmospheric Systems) sensor of Megha-Tropiques mission under Indo-French joint programme which will have almost similar frequencies as those of TRMM-TMI. Recently ANN has been recognized as being useful for retrieval operations in remote sensing of the ocean and atmosphere. The use of ANN in statistical estimation is often effective because it can simultaneously address nonlinear dependencies and complex statistical behaviour. It has been shown that a multilayer perceptron 4 with a single hidden layer and nonlinear activation function is capable of approximating any real valued continuous function, provided a sufficient number of units within the hidden layer exists 5 . Several studies have been conducted earlier, for retrieval of sea surface wind 6,7 , total precipi- table water 8–10 and integrated cloud liquid water 11–14 para- meters using mainly Special Sensor Microwave Imager (SSM/I) data and multiple regression approach. More re- cently, a methodology for the formulation of multi- parameter and multi-instrument retrieval from TRMM was developed by Obligis et al. 15 . The performance of ANN approach compared to the multiple regression ap- proach has recently been made 16 , for retrieval of TPW and CLW with the SSM/I observations. The present at- tempt is to evaluate the performance of ANN approach for retrieval of all the above-mentioned parameters using simulated database from radiative transfer model for TRMM-TMI radiometric channels. The ANN approach used here is based on a multilayer perceptron developed by Moreau et al. 17 . The main input database for radiative transfer simula- tions is from a set of ECMWF forecast fields, represent- ing tropical regions of the globe mainly during monsoon period over the Indian Ocean and Indian sub-continent. About 25,368 atmospheric profiles were used for the simulations through a radiative transfer model from Pri- gent et al. 18 that constitutes the vertical profiles of tem- perature, pressure, specific humidity and CLW defined on 31 vertical levels and the pressure, temperature and specific humidity at the surface and wind speed at 10 m height. The horizontal resolution of these data are 1.125° × 1.125° in latitude and longitude, which corresponds to a 125 km horizontal mesh at the equator. The model simu- lated brightness temperatures (TBs) and the correspond- ing ECMWF forecast fields of the main geophysical parameters of interest are used here as the input and output field vectors for designing the ANN architecture both for training and testing purposes. Given accurate and reliable input and output field vectors, geophysical retrieval algorithms can be developed using various approaches. As stated earlier, we have used the ANN approach, on the database using the radiative transfer simulations for TMI channels (10, 18.7, 21, 37 and 85 GHz) for both horizontal (H) and vertical (V) polari- zation except for a single-V polarization at 21 GHz, with the ECMWF fields of oceanic and atmospheric variables.