Application of a neural network to ERS-SAR images to retrieve pressure ridge spatial frequencies is presented. For an independent dataset, the rms-error between the retrieved and the true ridge frequency as determined by means of laser profiling was about 5 ridges per kilometer, or 30%. The network is trained with results from in-situ laser profiling of ridge distributions and coincident SAR backscatter properties. The study focusses on summer data from the Bellingshausen, Amundsen and Weddell Seas in Antarctica, which were gathered in February 1994 and 1997. Pressure ridge frequencies varied from 3 to 30 ridges per kilometer between different regions, thus providing a wide range of training and test data for the algorithm development.From ERS-SAR images covering the area of the laser flights with a time difference of a few days at maximum, histograms of the backscatter coefficient s0 were extracted. Statistical parameters (e.g. mean, standard deviation, tail-to-mean ratio) were calculated from these distributions and compared with the results of the laser flights. Generally, the mean backscatter increases with a growing ridge frequency, and the signal range becomes narrower. However, these correlations are only poor, and improved results are obtained when the statistical parameters are combined to train the neural network.