Automatic land cover classification of multi-resolution dualpol data using convolutional neural network (CNN)

Abstract

Synthetic Aperture Radar is an interesting topic of research for scientists & researchers as it is associated with polarimetric information which helps to detect surface & subsurface features of land, sea, and ice. Classical techniques include the use of polarimetric information to simplify SAR image interpretation and to classify it for various earth observation applications. The deep learning (DL) techniques like Convolutional Neural Network (CNN), extract useful information from an image (here dual polarimetric SAR dataset) about the land surface to segment or classify the dataset for various earth applications. In the current research paper convolutional neural network is used to automatically classify RISAT-1 dataset over the Mumbai region for land cover classification. Also impact of patch size variation was studied. In addition, the efficiency of the CNN model was tested using an approach similar to transfer learning approach on multi resolution images (different multilooked images) i.e. CNN was trained twice on different resolution images; one trained on coarser resolution and tested on comparatively higher resolution datasets and other vice versa. It was found that increasing the patch size for convolution classified the image more accurately but at the same time it smoothens the output image. Also, CNN model trained on low spatial resolution image predicted better the higher spatial resolution image as compared to the reverse scenario.