A novel multilevel pulse coupled neural networks architecture for objects recognition applied on ASI cosmo-skymed data

In this study a novel architecture of Pulse Coupled Neural Network based on a multilevel topology with interconnected layers is presented. The model is applied to the solution of a segmentation problem of SAR images for the identification of man-made structures on urban landscapes. Thanks to the multilayer architecture, the unsupervised model can deal with dual polarization SAR data, as well as with combination of ascending and descending acquisitions. Such approach mitigates the issues in detecting artificial targets when their orientation with respect the satellite line of sight reduces the object backscattering with respect to the one of background. An example of application to two COSMO-SkyMed STRIPMAP data, acquired by ascending and descending orbits respectively is provided.The proposed approach, owing to their ability of efficiently processing voluminous datasets, can be effectively coupled with machine learning and deep learning to refine or validate their results.