Terrain Radiometric Correction of SAR Images Based on Neural Network

Synthetic Aperture Radar (SAR) images suffer from significant radiation distortion due to the effect of terrain relief, which makes the further analyze and apply quantitatively of SAR data difficult. The radiometric correction approach based on neural network is done by building training data sets of ascending and descending SAR images after co-registration. Different network input features have a huge impact on the correction performance. The comparison of the terrain correction performance of SAR images in the Wuyi Mountain area with three different input schemes show that the optimal correction result can reduce the image variance by more than 770%, and the other two input schemes reduce the variance by more than 70%. The correction results in four subareas with distinct topographic indicate that the average correlation coefficient between the backscatter coefficient and local incidence angle fall from 0.880 before correction to 0.224, 0.134, and 0.0382 after correction respectively and the mean values of the image variance correspondingly decrease from 22.534 to 5.497, 5.153 and 4.371.