Csrs-Siat: A Benchmark Remote Sensing Dataset to Semantic-Enabled and Cross-Scales Scene Recognition

The deep learning has been widely used in scene recognition of remote sensing images. However, the accuracy of deep learning relays on the size of training dataset to the utmost. The remote sensing images have various spatial scales and semantics, which are not fully considered in the existing datasets. In this paper, a benchmark remote sensing dataset named as Cross-Scale Remote Sensing dataset of Shenzhen Institutes of Advanced Technology (CSRS-SIAT) is proposed, which has about 100 classes according to the land cover and land use field, and due to the cross-scale characteristics of remote sensing images, the experiments using traditional and state-of-the-art deep learning algorithms shows that there still need more efforts to achieve better results.