Deep learning based distributed scatterers acceleration approach: Distributed scatterers prediction Net

Distributed scatter (DS) interferometric synthetic aperture radar is a powerful technology for analyzing displacements of the earth's surface. Unfortunately, the preparatory step of DS pre-processing is enormously time consuming. The present research puts forward a deep learning-based approach called Distributed Scatterers Prediction Net (DSPN), that can reduce the computational load considerably. DSPN is a convolutional neural network, which generates DS candidate masks based on nine input layers. Masked pixels with low prospect of being DS are omitted during DS pre-processing. Tests on 6 different terrains in North Rhine-Westphalia and Sicily with Sentinel-1 data show that DSPN saves 11% to 87% computation time depending on the scene without significantly reducing coverage with information. Our experiments show that the proposed approach can effectively predict DS candidates and speeds up processing, indicating its potential for analyzing the big data of remote sensing. To the best of our knowledge, this is the first attempt to do a classification in DS candidates and non-DS candidates as a preparatory step to DS pre-processing.