Deep highway unit network for land cover type classification with GF-3 SAR imagery

Abstract

The fully polarized synthetic aperture radar (SAR) is an advanced earth observation system with day and night imaging capability, which can obtain rich information of terrain and has a wide range of applications in environmental protection, urban planning and resource investigation. As the first self-developed C-band multi-polarized SAR image, the acquisition of massive data and operational operation of Chinese SAR remote sensing has entered the era of big data. Under the era of remote sensing large data, however, SAR image interpretation is a great challenge for scientific applications. At present, big data-based intelligent methods such as computer vision technology have achieved great success. Deep learning such as deep highway unit networks has revolutionized the computer vision area. However, due to the characteristics of SAR microwave band imaging and phase coherence processing, SAR images are very different from ordinary optical images in terms of band, projection direction, data composition and so on. Therefore, deep learning can not be directly used for quad-pol SAR image classification. In this paper, deep learning is applied to land cover type classification with GF-3 quad-pol SAR imagery. A deep highway unit network is employed to automatically extract a hierarchic feature representation from the data, based on which the land cover type classification can be conducted. Our classification model is trained on limited training data from forest resource inventory and planning data, and tested on a Radarsat-2 quad-pol images, which is the image of the same area acquired at different times. We also employ the machine learning such as SVM, Random Forest on the same samples for comparison. The deep highway unit network trained by the GF-3 images, which can reduce speckle, fully excavate the regularity of SAR images in time and space.