Abstarct: Observations of earth such as on the oceans and on the land using the SAR ( synthetic aperture radar ) of the ERS ( earth resource satellite ) have wide range of applications. The ability of SAR to penetrate cloud over makes it particularly valuable in frequently clouded areas as tropics.SAR data can be used to georefer other satellite imagery to high precision and update the thematic maps more frequently and cost effective, due to its availability from weather conditions. In this paper we proposed and reports an investigation aimed at systematic preprocessing and a training strategy of the U net architecture for multi label classification that can classify multi number classes present SAR. In recent years many CNNs for SAR data classifications using deep learning have been proposed but most of failed to propose a perfect training strategy for target classes more than two using neural networks. Based on our results the proposed method can outcome the problem and can classifies more than two labelled classes.
Libraries and Software used: In this work we are using some software required for some preprocessing work like SNAP (sentinel application software), ARCGis 10.1 desktop,GIMP(GNU image manipulation program) and we are using the well know network in image segmentation that is U net architecture with strides(image patches) based classification. Tensorflow, Keras >= 1.0,OpenCV 3