OpenSARUrban: A Sentinel-1 SAR Image Dataset for Urban Interpretation

Abstract

Sentinel-1 mission provides a freely accessible opportunity for urban interpretation from synthetic aperture radar (SAR) images with specific resolution, which is of paramount importance for earth observation. In parallel, with the rapid development of advanced technologies, especially deep learning, it is urgently needed to construct a large-scale SAR dataset leading urban interpretation. This paper presents OpenSARUrban: a Sentinel-1 dataset dedicated to urban interpretation from SAR images, including a well-defined hierarchical annotation scheme, the data collection, the well-established procedures for dataset construction and organizations, the properties, visualizations, and applications of this dataset. Particularly, the OpenSARUrban provides 33358 image patches of SAR urban scene, covering 21 major cities of China, including 10 different categories, 4 kinds of formats, 2 kinds of polarization modes, and owning 5 essential properties: large-scale, diversity, specificity, reliability, and sustainability. These properties guarantee the achievable of several goals for OpenSARUrban. The first is to support urban target characterization. The second is to help develop applicable and advanced algorithms for Sentinel-1 urban target classification. The dataset visualization is implemented from the perspective of manifold to give an intuitive understanding. Besides a detailed description and visualization of the dataset, we present results of some benchmark algorithms, demonstrating that this dataset is practical and challenging. Notably, developing algorithms to enhance the classification performance on the whole dataset and considering the data imbalance are especially challenging.