Application of special-purpose artificial neural networks for speckle reduction in SAR images

Abstract

Synthetic aperture radar (SAR) is used extensively for remote-sensing applications due to its ability to operate under all weather conditions and provide high-resolution images. However, high-resolution images constructed from SAR data often suffer from speckle, which makes identification and classification of edges/boundaries a difficult task. Speckle noise is multiplicative in nature and is a result of constructive and destructive interference of signals from randomly distributed scatterers in a resolution cell illuminated by a coherent signal. Usually, speckle is reduced by incoherent averaging of high-resolution image pixels that degrade resolution. The principal goal in all speckle-reduction algorithms is to reduce speckle with minimum loss of resolution. In this investigation, we used specially trained and validated artificial neural networks (ANNs) for speckle reduction in images generated with a radar-depth sounder/imager and compared their performance to the conventional adaptive filtering and Speckle Reducing Anisotropic Diffusion (SRAD) algorithm. We show that by training different ANNs to reduce speckle noise at different levels of signal-to-noise ratio (SNR), rather than training one ANN to operate at all levels of SNR, improved performance in speckle reduction can be obtained. Real SAR images and synthetic noise are used in this research to compare the performance of the proposed ANN-based approaches with that obtained from conventional methods. This investigation shows that on combining the results from a set of properly trained and validated neural networks, the SNRs of the output images improve beyond those obtained from conventional approaches when the input SNRs are greater than or equal to 4 dB. For input SNRs greater than 0 dB, however, the ANNs provide better performance in edge preservation compared with conventional methods. We also found that once a set of ANNs is properly trained to reduce speckle from an image, these ANNs can be used in de-speckling other images without any further training. The merits and demerits of different configurations of the ANNs are studied to find useful speckle noise-tolerant ANN architectures.