Synthetic aperture radar (SAR) is a high-resolution remote sensing platform with all-weather capability. Traditional filter-based techniques are unsuitable for smoothing SAR images, but considerable success has been achieved using a CPU intensive, algorithmic noise removal process called simulated annealing. In order to reduce the CPU requirements of the despeckling process we have presented a solution based upon neural networks which are a form of adaptive filter. A variety of neural network architectures based on the multilayer perceptron and the vector quantizer network have been trained to learn the despeckling process. We have demonstrated that such a hybrid network can be successfully trained to perform speckle reduction of SAR images. The hybrid network benefits from reduced training and execution times compared to a single MLP, whilst maintaining a good performance.